-
-### Retrieval Accuracy
+### What's new in 0.9.31
-**coding-agent-life-v1** (in-house corpus, sandbox-reproducible)
+- **More focused recall:** project and agent scopes, saved-memory retrieval, explicit context budgets and full-source expansion.
+- **Less repeated work:** incremental session summaries skip unchanged observations; exports page through stored data.
+- **More native integrations:** installable Pi package, OMP and DeepSeek Harness plugins, and improved Cursor workspace attribution.
+- **Safer recovery:** compression, import validation and runtime recovery fixes, with status and activity visible in the viewer.
-| Adapter | P@5 | R@5 | Top-5 hit rate | p50 latency |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep baseline | 0.227 | 0.967 | 15 / 15 | 0 ms |
+[Release notes and contributor credits](https://github.com/rohitg00/agentmemory/releases/tag/v0.9.31).
-100% top-5 hit rate at the **P@5 math ceiling** for this corpus (0.240, see scorecard). Hybrid retrieves every gold session; grep misses 1 of 2 gold on the multi-session temporal query. Lift is **recall + temporal**, not aggregate precision. This benchmark is small and gold-sparse; the larger LongMemEval-S below differentiates better. Full per-type breakdown + correction note: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
+
-**LongMemEval-S** (ICLR 2025, 500 questions)
+Measured on 2026-10-28 with the 0.9.31 codebase. **Strict recall-all@5** requires every relevant source session in the first five results.
-| System | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| BM25-only fallback | 86.2% | 94.6% | 71.5% |
+| Dataset | BM25 component | grep baseline | Coverage |
+|---|---:|---:|---|
+| LongMemEval-S | **84.7% (398/470)** | 72.1% (339/470) | 500 questions per adapter; 30 unanswerable excluded from retrieval metrics |
+| coding-agent-life-v1 | 93.3% (14/15) | **100% (15/15)** | All 15 questions per adapter |
-
-
+All 1,030 attempts completed without execution failures. This tests supplied-session retrieval without a daemon, embeddings, compression or an answer-generating model. It does not measure full-system answer accuracy, release-to-release speedup or competitor performance.
-### Token Savings
+[Methodology, per-question receipts and reproduction](docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [Run the evaluations](eval/README.md).
-| Approach | Tokens/yr | Cost/yr |
-|---|---|---|
-| Paste full context | 19.5M+ | Impossible (exceeds window) |
-| LLM-summarized | ~650K | ~$500 |
-| **agentmemory** | **~170K** | **~$10** |
-| agentmemory + local embeddings | ~170K | **$0** |
-
-
-
-
-
-> Embedding model: `all-MiniLM-L6-v2` (local, free, no API key). Full reports: [`benchmark/LONGMEMEVAL.md`](benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](benchmark/QUALITY.md), [`benchmark/SCALE.md`](benchmark/SCALE.md). Competitor comparison: [`benchmark/COMPARISON.md`](benchmark/COMPARISON.md) covering agentmemory vs mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Reproduce locally:** [`eval/README.md`](eval/README.md), an adapter-pluggable harness for LongMemEval `_s` (public 500-Q) + `coding-agent-life-v1` (in-house 15-session corpus). Grep / vector / agentmemory adapters score side-by-side, NDJSON output, published scorecards land in [`docs/benchmarks/`](docs/benchmarks/).
-
-**Pairs with [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything), and [Graphify](https://github.com/safishamsi/graphify).** Code-graph indexing, multi-agent build pipelines, and broader knowledge graphs across docs / PDFs / images / videos. agentmemory remembers the work; those three projects light up the rest of the context layer. Recipes + question-routing table: [`docs/recipes/pairings.md`](docs/recipes/pairings.md).
+Historical [hybrid retrieval](benchmark/LONGMEMEVAL.md), [token use](benchmark/REAL-EMBEDDINGS.md) and [coding-session](docs/benchmarks/2026-05-20-coding-agent-life-v1.md) experiments use older versions or different metrics. Their scores are not current-release guarantees.
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Built-in (CLAUDE.md)
-
-
-
Type
-
Memory engine + MCP server
-
Memory layer API
-
Full agent runtime
-
Personal AI
-
Memory API + app
-
Team memory hub (LLM proxy)
-
Vector memory (OSS)
-
Memory engine (Oracle DB)
-
Memory system
-
Static file
-
-
-
Retrieval R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Self-reported
-
PersonaMem 76% (self-reported)
-
~96.6% (self-reported)
-
94.4% (self-reported)
-
N/A
-
N/A (grep)
-
-
-
Auto-capture
-
12 hooks (zero manual effort)
-
Manual add() calls
-
Agent self-edits
-
Manual
-
API-side extraction
-
Proxy interception (base-URL swap)
-
Manual
-
API extraction
-
Manual
-
Manual editing
-
-
-
Search
-
BM25 + Vector + Graph (RRF fusion)
-
Vector + Graph
-
Vector (archival)
-
Semantic
-
Vector + RAG
-
4 asset types (Chat / Skill / Wiki / CodeGraph)
-
Vector-only
-
Vector + semantic
-
Decay-weighted
-
Loads everything into context
-
-
-
Multi-agent
-
MCP + REST + leases + signals
-
API (no coordination)
-
Within Letta runtime only
-
No
-
No
-
Team roles + shared assets
-
No
-
Scoped only
-
Multi-agent shared
-
Per-agent files
-
-
-
Framework lock-in
-
None (any MCP client)
-
None
-
High (must use Letta)
-
Standalone
-
None
-
Proxy fronts every model call
-
None
-
Oracle Database
-
None
-
Per-agent format
-
-
-
External deps
-
No separate database service required
-
Qdrant / pgvector
-
Postgres + vector DB
-
Multiple
-
Managed cloud
-
Docker stack (Core + Hub + Proxy)
-
Vector store
-
Oracle AI Database
-
None
-
None
-
-
-
Memory lifecycle
-
4-tier consolidation + decay + auto-forget
-
Passive extraction
-
Agent-managed
-
Manual
-
Auto-forget
-
Manual review; auto-routing in progress
-
None
-
Not stated
-
Decay + consolidation
-
Manual pruning
-
-
-
Token efficiency
-
~1,900 tokens/session ($10/yr)
-
Varies by integration
-
Core memory in context
-
Varies
-
Cloud pricing
-
Not stated
-
No token budget
-
LLM-backed (varies)
-
Varies
-
22K+ tokens at 240 obs
-
-
-
Real-time viewer
-
Yes (port 3113)
-
Cloud dashboard
-
Cloud dashboard
-
Web UI
-
Cloud dashboard
-
Hub web UI
-
No
-
No
-
No
-
No
-
-
-
Self-hosted
-
Yes (default)
-
Optional
-
Optional
-
Yes
-
No (cloud-only)
-
Yes (Docker)
-
Yes
-
Yes (Oracle DB)
-
Yes
-
Yes
-
-
-
-Benchmark note: only agentmemory's R@5 is our own measured result (LongMemEval-S, reproducible from benchmark/COMPARISON.md). The mem0 and Letta figures are their published LoCoMo numbers (a different dataset); the MemPalace, supermemory, TencentDB (PersonaMem), and oracleagentmemory figures are vendor self-reported claims we have not independently reproduced (oracleagentmemory's run used GPT-5.5 against an Oracle AI Database). Shown side by side for ballpark only, not a head-to-head on identical data. Star counts are approximate and drift over time.
+Choose by workload and deployment. These projects have evolved beyond a simple local-versus-cloud split.
-**Newer entrants** worth knowing, compared in depth in [`benchmark/COMPARISON.md`](benchmark/COMPARISON.md):
-
-| System | ⭐ | Angle |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporal knowledge graph; strongest published temporal-query results (LongMemEval 63.8%), but graph builds asynchronously so fresh facts can lag |
-| Cognee | 30K | Document-to-knowledge-graph ingestion, Python-only, built for structured entity extraction rather than session capture |
+| Project | Best starting point for |
+|---|---|
+| **AgentMemory** | Shared local memory for existing coding agents, automatic capture, scoped recall and a live viewer |
+| [Mem0](https://docs.mem0.ai/introduction) | Memory APIs, hosted or self-hosted, with SDKs and agent plugins |
+| [Supermemory](https://supermemory.ai/docs/self-hosting/overview) | Memory and document context through a hosted API or local server |
+| [Zep / Graphiti](https://github.com/getzep/graphiti) | Enterprise context and temporal knowledge graphs |
+| [Letta](https://docs.letta.com/) | Building stateful agents with their own managed memory |
+| [Cognee](https://github.com/topoteretes/cognee) | Turning documents, code and conversations into searchable knowledge |
+| [claude-mem](https://github.com/thedotmack/claude-mem) | Capturing and summarizing coding sessions across supported agents |
-AgentMemory combines coding-agent capture, searchable project history and a local viewer, with keyless operation available from the first run.
+[Detailed comparison and sources](benchmark/COMPARISON.md), checked 2026-10-28. Deployment options, provider costs and integrations vary; scores from different datasets and metrics are not a head-to-head ranking.
---
@@ -787,9 +607,9 @@ Full guide: [`integrations/hermes/`](integrations/hermes/)
Start the memory server: `npx -y @agentmemory/agentmemory@latest`
-#### Native skills via `npx skills add` (50+ agents)
+#### Native skills via `npx skills add`
-agentmemory ships 17 skills in the Claude-Code-style `/SKILL.md` format: 9 invocable action skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) and 8 reference skills the agent loads on demand (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). The reference skills carry data tables generated from source, so they never drift. The [`skills`](https://npmjs.com/package/skills) CLI by vercel-labs auto-installs them into the calling agent's native skill directory across 50+ agents (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, and more):
+agentmemory ships 17 skills in the Claude-Code-style `/SKILL.md` format: 9 invocable action skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) and 8 reference skills the agent loads on demand (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). The reference skills carry data tables generated from source, so they never drift. The [`skills`](https://npmjs.com/package/skills) CLI by vercel-labs auto-installs them into the calling agent's native skill directory across supported agents (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, and more):
```bash
npx skills add rohitg00/agentmemory -y # auto-detects the calling agent
@@ -845,7 +665,7 @@ The agentmemory entry is the **same MCP server block** across every host that us
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` installs MCP and capture hooks in the shared customization directory. See [Antigravity setup and limits](docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` uses the same MCP and hook configuration as current IDE versions. Existing installations should refresh with `--force`; see the [upgrade notes](docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` writes the user-level config. Workspace overrides go in `.kiro/settings/mcp.json` next to your code. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` writes the standard `mcpServers` block. Warp also auto-discovers skills from `.claude/skills/`; once the Claude Code plugin is installed the 8 agentmemory skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) appear natively in Warp's slash-command palette. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` writes the standard `mcpServers` block. Warp also auto-discovers skills from `.claude/skills/`; once the Claude Code plugin is installed the 9 agentmemory skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) appear natively in Warp's slash-command palette. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` writes the standard `mcpServers` block. VS Code extension users: paste the same block via Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (preferred) or `config.json` (legacy) | `agentmemory connect continue` creates `config.yaml` from scratch when neither exists, or modifies existing `config.json`. **If you already have `config.yaml`** the adapter prints the exact block to paste under `mcpServers:`; it won't silently rewrite your yaml so your existing comments and anchors are preserved. Continue uses array form (not object) for `mcpServers`. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` writes under `context_servers` (Zed's key, NOT `mcpServers`). Remote MCP servers can be wired via `{"url": "..."}` instead. |
@@ -938,15 +758,15 @@ Session 2: "Now add rate limiting"
### vs built-in agent memory
-Every AI coding agent ships with built-in memory: Claude Code has `MEMORY.md`, Cursor has notepads, Cline has memory bank. These work like sticky notes. agentmemory is the searchable database behind the sticky notes.
+Agent instructions and memory files remain useful for stable project guidance. AgentMemory adds searchable session history, source-linked memories and coordination across connected agents.
| | Built-in (CLAUDE.md) | agentmemory |
|---|---|---|
| Upkeep | Manually curated instructions | Indexed session history and memories |
| Search | Loads everything into context | BM25 + vector + graph (top-K only) |
-| Token cost | 22K+ at 240 observations | ~1,900 tokens (92% less) |
+| Context budget | Depends on the files included | Configurable recall budget with source expansion |
| Cross-agent | Per-agent files | MCP + REST (any agent) |
-| Coordination | None | Leases, signals, actions, routines |
+| Coordination | Depends on the agent | Leases, signals, actions, routines |
| Observability | Read files manually | Real-time viewer on :3113 |
---
@@ -957,7 +777,7 @@ Every AI coding agent ships with built-in memory: Claude Code has `MEMORY.md`, C
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID, durable capture inbox, offline spool when unavailable
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -966,7 +786,7 @@ PostToolUse hook fires
-> Index in BM25, plus vectors when enabled
Stop / SessionEnd hook fires
- -> Summarize session
+ -> Incremental summary when new observations are available
-> Knowledge graph extraction (if GRAPH_EXTRACTION_ENABLED=true)
-> Slot reflection (if AGENTMEMORY_SLOTS=true and AGENTMEMORY_REFLECT=true)
@@ -1046,7 +866,7 @@ Fused with Reciprocal Rank Fusion (RRF, k=60) and session-diversified (max 3 res
When a vector index is populated, `mem::search` (behind `memory_recall`) uses the hybrid BM25 + vector ranker. Without embeddings it uses BM25. `smart-search` can additionally fuse structural graph matches when graph data exists, including in keyless mode. Lesson recall runs on a dedicated in-memory BM25 index instead of scanning the whole corpus per query. Superseded memory versions are excluded from every recall path; the version chain keeps their history.
-Vectors survive a crash or force-kill. The vector index is saved in buckets at most every `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minutes). Every vector added or removed in between is also written right away to a small pending log in the state store, and the next start replays it without calling the embedding provider. Each successful save empties the log. Documents that still have no vector after the replay are re-embedded in the background in batches of `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) until none are left, and a backfill that is stopped continues at the next start. `/agentmemory/status` and the viewer show the pending log size and the backfill state. Keyless installs write nothing.
+Vectors survive a crash or force-kill. The vector index is saved in buckets at most every `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minutes). Every vector added or removed in between is also written right away to a small pending log in the state store, and the next start replays it without calling the embedding provider. Each successful save empties the log. For older documents without vectors, `AGENTMEMORY_VECTOR_BACKFILL=all` opts into a bounded backfill at startup, capped by `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500). Keep the opt-in set across restarts to process more documents. See [source and vector recovery](docs/reliability/source-and-vector-recovery.md). `/agentmemory/status` and the viewer show the pending log size and the backfill state. Keyless installs write nothing.
BM25 tokenizes Greek, Cyrillic, Hebrew, Arabic, and accented Latin out of the box. For Chinese / Japanese / Korean memories, install the optional segmenters (`npm install @node-rs/jieba tiny-segmenter`) to split CJK runs into word-level tokens; without them, agentmemory soft-falls to whole-run tokenization and prints a one-time hint on stderr.
@@ -1062,9 +882,9 @@ The normal npm install includes the optional `@huggingface/transformers` runtime
| Provider | Model | Cost | Notes |
|---|---|---|---|
-| **Local (recommended opt-in)** | `all-MiniLM-L6-v2` | Free | On-device after the first model download, +8pp recall over BM25-only |
+| **Local (recommended opt-in)** | `all-MiniLM-L6-v2` | Free | On-device after the first model download |
| Gemini | `gemini-embedding-001` | Free tier | 100+ languages, 768/1536/3072 dims (MRL), 2048-token input. Replaces `text-embedding-004` ([deprecated, shutdown Jan 14, 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Highest quality |
+| OpenAI | `text-embedding-3-small` | [Provider pricing](https://openai.com/api/pricing/) | Remote embeddings |
| Voyage AI | `voyage-code-3` | Paid | Optimized for code |
| Cohere | `embed-english-v3.0` | Free trial | General purpose |
| OpenRouter | Any model | Varies | Multi-model proxy |
diff --git a/READMEs/README.ar-SA.md b/READMEs/README.ar-SA.md
index c353c0337e..e3e0015ac6 100644
--- a/READMEs/README.ar-SA.md
+++ b/READMEs/README.ar-SA.md
@@ -68,8 +68,6 @@
-
-### دقة الاسترجاع
-
-**coding-agent-life-v1** (مجموعة بيانات داخلية، قابلة لإعادة الإنتاج في بيئة معزولة)
-
-| المحوّل | P@5 | R@5 | معدل الإصابة ضمن أفضل 5 | زمن الاستجابة p50 |
-|---|---|---|---|---|
-| **agentmemory الهجين** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| خط أساس grep | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-معدل إصابة 100% ضمن أفضل 5 عند **سقف P@5 الرياضي** لهذه المجموعة (0.240، راجع بطاقة النتائج). يسترجع النظام الهجين كل جلسة مرجعية (gold)؛ بينما يفوّت grep جلسة واحدة من جلستين مرجعيتين في الاستعلام الزمني متعدد الجلسات. التحسّن هنا في **الاسترجاع + البعد الزمني**، لا في الدقة الكلية. هذا المقياس صغير وقليل الجلسات المرجعية؛ ويميّز المقياس الأكبر LongMemEval-S أدناه بشكل أفضل. التفصيل الكامل حسب النوع وملاحظة التصحيح: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025، 500 سؤال)
+نتائج الاسترجاع الحالية، وتعريف المقاييس، وخطوات إعادة الاختبار:
-| النظام | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| BM25 فقط (خيار احتياطي) | 86.2% | 94.6% | 71.5% |
-
-
-
-> نموذج التضمين: `all-MiniLM-L6-v2` (محلي، مجاني، بلا مفتاح API). التقارير الكاملة: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md)، [`benchmark/QUALITY.md`](../benchmark/QUALITY.md)، [`benchmark/SCALE.md`](../benchmark/SCALE.md). مقارنة بالمنافسين: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) وتشمل agentmemory في مواجهة mem0 وLetta وKhoj وsupermemory وTencentDB Agent Memory وMemPalace وZep/Graphiti وCognee وHippo.
-
-**إعادة الإنتاج محليًا:** [`eval/README.md`](../eval/README.md)، وهو إطار اختبار (harness) قابل لتوصيل المحوّلات لـ LongMemEval `_s` (500 سؤال علنية) + `coding-agent-life-v1` (مجموعة داخلية من 15 جلسة). تُقيَّم محوّلات grep / المتجهات / agentmemory جنبًا إلى جنب، بمخرجات NDJSON، وتُنشر بطاقات النتائج في [`docs/benchmarks/`](../docs/benchmarks/).
-
-**يتكامل مع [codegraph](https://github.com/colbymchenry/codegraph) و[Understand Anything](https://github.com/Lum1104/Understand-Anything) و[Graphify](https://github.com/safishamsi/graphify).** فهرسة الرسم البياني للشفرة (code-graph)، وخطوط بناء متعددة الوكلاء، ورسوم بيانية معرفية أوسع تمتد عبر المستندات/ملفات PDF/الصور/الفيديوهات. يتذكّر agentmemory العمل؛ وتُضيء هذه المشاريع الثلاثة بقية طبقة السياق. الوصفات وجدول توجيه الأسئلة: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
مدمج (CLAUDE.md)
-
-
-
النوع
-
محرك ذاكرة + خادم MCP
-
واجهة برمجية لطبقة الذاكرة
-
بيئة تشغيل كاملة للوكيل
-
ذكاء اصطناعي شخصي
-
واجهة برمجية للذاكرة + تطبيق
-
محور ذاكرة جماعي (وكيل LLM وسيط)
-
ذاكرة متجهية (مفتوحة المصدر)
-
محرك ذاكرة (قاعدة بيانات Oracle)
-
نظام ذاكرة
-
ملف ثابت
-
-
-
دقة الاسترجاع R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
غير متوفر
-
مُعلَن من الشركة نفسها
-
PersonaMem 76% (مُعلَن من الشركة نفسها)
-
~96.6% (مُعلَن من الشركة نفسها)
-
94.4% (مُعلَن من الشركة نفسها)
-
غير متوفر
-
غير متوفر (grep)
-
-
-
الالتقاط التلقائي
-
12 خطافًا (بلا أي جهد يدوي)
-
استدعاءات add() يدوية
-
تعديلات ذاتية من الوكيل
-
يدوي
-
استخراج من جانب الواجهة البرمجية
-
اعتراض عبر وكيل وسيط (باستبدال base-URL)
-
يدوي
-
استخراج عبر API
-
يدوي
-
تعديل يدوي
-
-
-
البحث
-
BM25 + المتجهات + الرسم البياني (دمج RRF)
-
المتجهات + الرسم البياني
-
متجهات (أرشيفية)
-
دلالي
-
متجهات + RAG
-
4 أنواع من الأصول (محادثة / مهارة / Wiki / CodeGraph)
-
متجهات فقط
-
متجهات + دلالي
-
مرجَّح بالتلاشي
-
يحمّل كل شيء إلى السياق
-
-
-
تعدد الوكلاء
-
MCP + REST + حجوزات (leases) + إشارات
-
API (بلا تنسيق)
-
داخل بيئة تشغيل Letta فقط
-
لا
-
لا
-
أدوار جماعية + أصول مشتركة
-
لا
-
محدود النطاق فقط
-
مشترك بين وكلاء متعددين
-
ملفات لكل وكيل على حدة
-
-
-
الارتهان لإطار عمل محدد
-
لا يوجد (أي عميل MCP)
-
لا يوجد
-
مرتفع (يجب استخدام Letta)
-
مستقل
-
لا يوجد
-
وكيل وسيط يتقدّم كل استدعاء للنموذج
-
لا يوجد
-
Oracle Database
-
لا يوجد
-
تنسيق خاص بكل وكيل
-
-
-
الاعتماديات الخارجية
-
لا يتطلب خدمة قاعدة بيانات منفصلة
-
Qdrant / pgvector
-
Postgres + قاعدة بيانات متجهية
-
متعددة
-
سحابة مُدارة
-
حزمة Docker (Core + Hub + Proxy)
-
مخزن متجهات
-
Oracle AI Database
-
لا يوجد
-
لا يوجد
-
-
-
دورة حياة الذاكرة
-
توحيد رباعي المستويات + تلاشي + نسيان تلقائي
-
استخراج سلبي (passive)
-
يديرها الوكيل
-
يدوي
-
نسيان تلقائي
-
مراجعة يدوية؛ والتوجيه التلقائي قيد التطوير
-
لا يوجد
-
غير مُحدَّد
-
تلاشي + توحيد
-
تقليم يدوي
-
-
-
كفاءة استخدام الرموز
-
~1,900 رمز/جلسة (10$/سنويًا)
-
يتفاوت حسب التكامل
-
الذاكرة الأساسية داخل السياق
-
متفاوت
-
تسعير سحابي
-
غير مُحدَّد
-
بلا حد لعدد الرموز
-
معتمد على LLM (متفاوت)
-
متفاوت
-
أكثر من 22K رمز عند 240 ملاحظة
-
-
-
العارض اللحظي
-
نعم (المنفذ 3113)
-
لوحة تحكم سحابية
-
لوحة تحكم سحابية
-
واجهة ويب
-
لوحة تحكم سحابية
-
واجهة ويب للمحور (Hub)
-
لا
-
لا
-
لا
-
لا
-
-
-
استضافة ذاتية
-
نعم (افتراضيًا)
-
اختياري
-
اختياري
-
نعم
-
لا (سحابي فقط)
-
نعم (Docker)
-
نعم
-
نعم (Oracle DB)
-
نعم
-
نعم
-
-
-
-ملاحظة على المقياس: نتيجة R@5 الخاصة بـ agentmemory فقط هي قياسنا الفعلي (LongMemEval-S، قابلة لإعادة الإنتاج من benchmark/COMPARISON.md). أما أرقام mem0 وLetta فهي أرقامهما المنشورة على LoCoMo (مجموعة بيانات مختلفة)؛ وأرقام MemPalace وsupermemory وTencentDB (PersonaMem) وoracleagentmemory هي ادعاءات مُعلَنة من الشركات نفسها لم نُعِد إنتاجها بشكل مستقل (استخدمت عملية تشغيل oracleagentmemory نموذج GPT-5.5 مقابل قاعدة بيانات Oracle AI Database). عُرضت جنبًا إلى جنب للمقارنة التقريبية فقط، لا كمواجهة مباشرة على بيانات متطابقة. أعداد النجوم تقريبية وتتغير مع الوقت.
-
-**داخلون جدد** تستحق المعرفة، مع مقارنة تفصيلية في [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| النظام | ⭐ | الزاوية |
-|--------|---|-------|
-| Zep / Graphiti | 30K | رسم بياني معرفي زمني؛ يملك أقوى نتائج منشورة للاستعلامات الزمنية (LongMemEval 63.8%)، لكن بناء الرسم البياني يتم بشكل غير متزامن، فقد تتأخر الحقائق الجديدة |
-| Cognee | 30K | استيعاب يحوّل المستندات إلى رسم بياني معرفي، مخصص لـ Python فقط، ومبني لاستخراج الكيانات البنيوية لا لالتقاط الجلسات |
+مقارنة محدثة بين المنتجات وخيارات النشر، مع المصادر:
-يجمع AgentMemory التقاط الجلسات وسجل المشروع القابل للبحث والعارض المحلي، مع العمل دون مفتاح منذ التشغيل الأول.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -928,7 +720,6 @@ npx -y @agentmemory/mcp
|---|---|---|
| الصيانة | تعليمات تُدار يدويًا | سجل جلسات وذكريات مفهرسة |
| البحث | يحمّل كل شيء إلى السياق | BM25 + متجهات + رسم بياني (أفضل K فقط) |
-| تكلفة الرموز | أكثر من 22K عند 240 ملاحظة | ~1,900 رمز (أقل بنسبة 92%) |
| بين الوكلاء | ملفات لكل وكيل | MCP + REST (أي وكيل) |
| التنسيق | لا يوجد | حجوزات، وإشارات، وإجراءات، ومسارات عمل (routines) |
| قابلية الرصد | قراءة الملفات يدويًا | عارض لحظي على :3113 |
@@ -941,7 +732,7 @@ npx -y @agentmemory/mcp
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1026,7 +817,7 @@ SessionStart hook fires
عندما يكون فهرس المتجهات مُعبَّأً، يستخدم `mem::search` (الذي يقف خلف `memory_recall`) المُصنِّف الهجين BM25 + المتجهات. وبدون تضمينات يستخدم BM25. ويمكن لـ `smart-search` أيضًا دمج تطابقات الرسم البياني البنيوية عندما تتوفر بيانات رسم بياني، حتى في وضع عدم استخدام المفاتيح. ويعمل استرجاع الدروس المستفادة على فهرس BM25 مخصص في الذاكرة بدلًا من مسح المجموعة الكاملة في كل استعلام. وتُستبعد نسخ الذاكرة المُستبدَلة من كل مسار استرجاع؛ وتحتفظ سلسلة الإصدارات بتاريخها.
-تنجو المتجهات من الانهيار أو الإيقاف القسري. ويُحفظ فهرس المتجهات في حُزَم مرة واحدة على الأكثر كل `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 دقائق). ويُكتب أيضًا كل متجه يُضاف أو يُحذف بين ذلك فورًا في سجل انتظار صغير في مخزن الحالة، ويُعاد تشغيله عند البدء التالي دون استدعاء مزوّد التضمين. ويُفرَّغ السجل عند كل حفظ ناجح. وتُعاد تضمين المستندات التي لا تزال بلا متجه بعد إعادة التشغيل في الخلفية على دفعات بحجم `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) إلى أن لا يبقى منها شيء، ويستكمل أي ملء رجعي (backfill) متوقف في البدء التالي. يُظهر `/agentmemory/status` والعارض حجم سجل الانتظار وحالة الملء الرجعي. ولا تكتب التثبيتات بلا مفاتيح أي شيء.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
يُجزِّئ BM25 إلى رموز (tokens) اليونانية والسيريلية والعبرية والعربية واللاتينية المُشكَّلة جاهزًا دون أي إعداد. وبالنسبة لذكريات الصينية/اليابانية/الكورية، ثبّت المُجزِّئات (segmenters) الاختيارية (`npm install @node-rs/jieba tiny-segmenter`) لتقسيم سلاسل CJK إلى رموز على مستوى الكلمة؛ وبدونها، يتراجع agentmemory بسلاسة إلى تجزيء السلسلة كاملةً ويطبع تلميحًا لمرة واحدة على stderr.
@@ -1042,9 +833,9 @@ EMBEDDING_PROVIDER=local
| المزوّد | النموذج | التكلفة | ملاحظات |
|---|---|---|---|
-| **محلي (مُوصى بتفعيله)** | `all-MiniLM-L6-v2` | مجاني | على الجهاز نفسه بعد أول تنزيل للنموذج، وبزيادة 8 نقاط نسبة مئوية في الاسترجاع عن BM25 فقط |
+| **محلي (مُوصى بتفعيله)** | `all-MiniLM-L6-v2` | مجاني | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | مستوى مجاني | أكثر من 100 لغة، أبعاد 768/1536/3072 (MRL)، ومدخل حتى 2048 رمزًا. يحل محل `text-embedding-004` ([مُستبعَد، يُغلق في 14 يناير 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | 0.02$/مليون | أعلى جودة |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | مدفوع | مُحسَّن للشفرة |
| Cohere | `embed-english-v3.0` | نسخة تجريبية مجانية | للاستخدام العام |
| OpenRouter | أي نموذج | متفاوت | وكيل متعدد النماذج |
diff --git a/READMEs/README.bn-BD.md b/READMEs/README.bn-BD.md
index 4346dd2b65..f484461278 100644
--- a/READMEs/README.bn-BD.md
+++ b/READMEs/README.bn-BD.md
@@ -66,8 +66,6 @@
Manuální revize; automatické směrování se připravuje
-
Žádný
-
Neuvedeno
-
Rozpad + konsolidace
-
Manuální prořezávání
-
-
-
Efektivita tokenů
-
~1,900 tokenů/session ($10/rok)
-
Liší se podle integrace
-
Klíčová paměť v kontextu
-
Liší se
-
Cloudové ceny
-
Neuvedeno
-
Žádný rozpočet tokenů
-
Založeno na LLM (liší se)
-
Liší se
-
22K+ tokenů při 240 obs.
-
-
-
Viewer v reálném čase
-
Ano (port 3113)
-
Cloudový dashboard
-
Cloudový dashboard
-
Webové UI
-
Cloudový dashboard
-
Webové UI hubu
-
Ne
-
Ne
-
Ne
-
Ne
-
-
-
Self-hosted
-
Ano (výchozí)
-
Volitelné
-
Volitelné
-
Ano
-
Ne (pouze cloud)
-
Ano (Docker)
-
Ano
-
Ano (Oracle DB)
-
Ano
-
Ano
-
-
-
-Poznámka k benchmarku: pouze R@5 pro agentmemory je náš vlastní naměřený výsledek (LongMemEval-S, reprodukovatelný z benchmark/COMPARISON.md). Čísla pro mem0 a Letta jsou jejich publikované hodnoty LoCoMo (jiný dataset); čísla pro MemPalace, supermemory, TencentDB (PersonaMem) a oracleagentmemory jsou tvrzení výrobců, která jsme nezávisle nereprodukovali (běh oracleagentmemory použil GPT-5.5 proti Oracle AI Database). Uvedeno vedle sebe jen pro orientační srovnání, nejde o přímé porovnání na identických datech. Počty hvězd jsou orientační a v čase se mění.
-
-**Novější konkurenti**, o kterých má smysl vědět, podrobně srovnáni v [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Systém | ⭐ | Přístup |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporální znalostní graf; nejsilnější publikované výsledky pro temporální dotazy (LongMemEval 63.8%), ale graf se buduje asynchronně, takže čerstvá fakta mohou zpožďovat |
-| Cognee | 30K | Načítání dokumentů do znalostního grafu, pouze Python, navrženo pro strukturovanou extrakci entit, nikoli pro zachytávání sessions |
+Aktuální srovnání produktů a možností nasazení se zdroji:
-AgentMemory spojuje zachycení sessions, prohledávatelnou historii projektu a lokální viewer; bez klíče funguje od prvního spuštění.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -786,7 +578,7 @@ Verify with `curl http://localhost:3111/agentmemory/health`. Open http://localho
Spusťte server paměti: `npx -y @agentmemory/agentmemory@latest`
-#### Nativní skills přes `npx skills add` (50+ agentů)
+#### Nativní skills přes `npx skills add`
agentmemory dodává 17 skills ve formátu `/SKILL.md` ve stylu Claude Code: 9 vyvolatelných akčních skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) a 8 referenčních skills, které agent načítá podle potřeby (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Referenční skills obsahují datové tabulky generované ze zdroje, takže se nikdy neliší. CLI [`skills`](https://npmjs.com/package/skills) od vercel-labs je automaticky instaluje do nativního adresáře skills volajícího agenta napříč 50+ agenty (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf a dalšími):
@@ -844,7 +636,7 @@ Položka agentmemory je **stejný blok serveru MCP** na všech hostitelích, kte
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` nainstaluje MCP a hooky pro zachycení do sdíleného adresáře přizpůsobení. Viz [nastavení a limity Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` používá stejnou konfiguraci MCP a hooků jako aktuální verze IDE. Stávající instalace by měly provést refresh pomocí `--force`; viz [poznámky k upgradu](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` zapíše konfiguraci na úrovni uživatele. Přepisy na úrovni workspace patří do `.kiro/settings/mcp.json` vedle vašeho kódu. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` zapíše standardní blok `mcpServers`. Warp také automaticky vyhledává skills z `.claude/skills/`; po instalaci pluginu Claude Code se 8 skills agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) objeví nativně v paletě slash příkazů Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` zapíše standardní blok `mcpServers`. Warp také automaticky vyhledává skills z `.claude/skills/`; po instalaci pluginu Claude Code se 9 skills agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) objeví nativně v paletě slash příkazů Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` zapíše standardní blok `mcpServers`. Uživatelé rozšíření VS Code: vložte stejný blok přes Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (preferované) nebo `config.json` (starší) | `agentmemory connect continue` vytvoří `config.yaml`, pokud konfigurace chybí, nebo upraví existující `config.json`. U existujícího YAML vypíše blok k vložení pod `mcpServers:`, aby zachoval komentáře a kotvy. Continue používá pole, nikoli objekt. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` zapíše pod `context_servers` (klíč Zed, NE `mcpServers`). Vzdálené servery MCP lze zapojit místo toho přes `{"url": "..."}`. |
@@ -927,7 +719,6 @@ Každý AI kódovací agent se dodává s vestavěnou pamětí: Claude Code má
|---|---|---|
| Údržba | Ručně spravované instrukce | Indexovaná historie sessions a vzpomínky |
| Vyhledávání | načte vše do kontextu | BM25 + vektor + graf (pouze top-K) |
-| Náklady na tokeny | 22K+ při 240 pozorováních | ~1,900 tokenů (o 92% méně) |
| Napříč agenty | soubory po agentech | MCP + REST (jakýkoli agent) |
| Koordinace | žádná | leases, signály, akce, routiny |
| Observabilita | manuální čtení souborů | viewer v reálném čase na :3113 |
@@ -940,7 +731,7 @@ Každý AI kódovací agent se dodává s vestavěnou pamětí: Claude Code má
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1025,7 +816,7 @@ Sloučeno pomocí Reciprocal Rank Fusion (RRF, k=60) a diverzifikováno po sessi
Když je vektorový index naplněný, `mem::search` (za `memory_recall`) používá hybridní řadič BM25 + vektor. Bez embeddingů používá BM25. `smart-search` může navíc sloučit strukturální shody z grafu, pokud grafová data existují, včetně režimu bez klíče. Vyhledávání lessons běží na vyhrazeném in-memory indexu BM25 místo skenování celého korpusu při každém dotazu. Nahrazené verze paměti jsou vyloučeny z každé cesty vyhledávání; řetězec verzí uchovává jejich historii.
-Vektory přežijí pád nebo násilné vypnutí. Vektorový index se ukládá v dávkách nejvýše jednou za `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minut). Každý vektor přidaný nebo odstraněný mezitím se navíc okamžitě zapíše do malého čekajícího logu ve state store a při dalším startu se přehraje bez volání poskytovatele embeddingů. Každé úspěšné uložení log vyprázdní. Dokumenty, které po přehrání stále nemají vektor, se na pozadí znovu embedují v dávkách po `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500), dokud žádné nezbydou, a zastavený backfill pokračuje při dalším startu. `/agentmemory/status` a viewer zobrazují velikost čekajícího logu a stav backfillu. Instalace bez klíče nic nezapisují.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 standardně tokenizuje řečtinu, cyrilici, hebrejštinu, arabštinu a akcentovanou latinku. Pro paměti v čínštině / japonštině / korejštině nainstalujte volitelné segmentery (`npm install @node-rs/jieba tiny-segmenter`), aby se běhy CJK rozdělily na tokeny na úrovni slov; bez nich agentmemory měkce spadne na tokenizaci celého běhu a jednou vypíše nápovědu na stderr.
@@ -1041,9 +832,9 @@ Běžná instalace npm obsahuje volitelný runtime `@huggingface/transformers`.
| Poskytovatel | Model | Náklady | Poznámky |
|---|---|---|---|
-| **Lokální (doporučeno jako opt-in)** | `all-MiniLM-L6-v2` | zdarma | na zařízení po prvním stažení modelu, +8pp recall oproti samotnému BM25 |
+| **Lokální (doporučeno jako opt-in)** | `all-MiniLM-L6-v2` | zdarma | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | bezplatný tier | 100+ jazyků, rozměry 768/1536/3072 (MRL), vstup 2048 tokenů. Nahrazuje `text-embedding-004` ([deprecated, ukončení 14. 1. 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | nejvyšší kvalita |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | placené | optimalizováno pro kód |
| Cohere | `embed-english-v3.0` | bezplatná zkouška | univerzální použití |
| OpenRouter | jakýkoli model | liší se | proxy pro více modelů |
diff --git a/READMEs/README.da-DK.md b/READMEs/README.da-DK.md
index 13dced179d..5eac7f2f04 100644
--- a/READMEs/README.da-DK.md
+++ b/READMEs/README.da-DK.md
@@ -66,8 +66,6 @@
-
-### Retrieval-nøjagtighed
-
-**coding-agent-life-v1** (internt korpus, reproducerbart i sandbox)
-
-| Adapter | P@5 | R@5 | Top-5-hitrate | p50-latens |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep-baseline | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-100% top-5-hitrate ved **P@5-matematikkens loft** for dette korpus (0.240, se scorecard). Hybrid finder hver gylden session; grep misser 1 af 2 gyldne i den multi-session-temporale forespørgsel. Forbedringen er **recall + temporal**, ikke samlet precision. Dette benchmark er lille og har få gyldne eksempler; det større LongMemEval-S nedenfor differentierer bedre. Fuld opdeling pr. type + korrektionsnote: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 spørgsmål)
+Aktuelle søgeresultater, definitioner af målinger og trin til gentagelse:
-| System | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| Kun BM25-fallback | 86.2% | 94.6% | 71.5% |
-
-
-
-> Embedding-model: `all-MiniLM-L6-v2` (lokal, gratis, ingen API-nøgle). Fulde rapporter: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Konkurrentsammenligning: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md), der dækker agentmemory vs. mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Reproducér lokalt:** [`eval/README.md`](../eval/README.md), en adapter-plugbar harness til LongMemEval `_s` (offentligt 500-spørgsmåls-sæt) + `coding-agent-life-v1` (internt korpus med 15 sessioner). Grep-, vektor- og agentmemory-adaptere scorer side om side, NDJSON-output, publicerede scorecards lander i [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Fungerer godt sammen med [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) og [Graphify](https://github.com/safishamsi/graphify).** Code-graph-indeksering, multi-agent build-pipelines og bredere videngrafer tværs af dokumenter/PDF'er/billeder/videoer. agentmemory husker arbejdet; disse tre projekter oplyser resten af kontekstlaget. Opskrifter + spørgsmålsrouting-tabel: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-Benchmark-note: kun agentmemorys R@5 er vores eget målte resultat (LongMemEval-S, reproducerbart fra benchmark/COMPARISON.md). Tallene for mem0 og Letta er deres publicerede LoCoMo-tal (et andet datasæt); tallene for MemPalace, supermemory, TencentDB (PersonaMem) og oracleagentmemory er leverandørens selvrapporterede påstande, som vi ikke selv har reproduceret (oracleagentmemorys kørsel brugte GPT-5.5 mod en Oracle AI Database). Vist side om side kun som et groft overblik, ikke en direkte sammenligning på identiske data. Stjerneantal er cirka og ændrer sig over tid.
-
-**Nyere aktører**, som er værd at kende, sammenlignet i detaljer i [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| System | ⭐ | Vinkel |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporal videngraf; stærkeste publicerede resultater for temporale forespørgsler (LongMemEval 63.8%), men grafen bygges asynkront, så nye fakta kan halte |
-| Cognee | 30K | Dokument-til-videngraf-indlæsning, kun Python, bygget til struktureret entitetsudtrækning snarere end sessionsindsamling |
+Opdateret sammenligning af produkter og driftsmuligheder med kilder:
-AgentMemory kombinerer registrering fra kodningsagenter, søgbar projekthistorik og en lokal viewer. Du kan arbejde uden nøgle fra første start.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -779,7 +571,7 @@ Fuld guide: [`integrations/hermes/`](../integrations/hermes/)
Start hukommelsesserveren: `npx -y @agentmemory/agentmemory@latest`
-#### Native skills via `npx skills add` (50+ agenter)
+#### Native skills via `npx skills add`
agentmemory leveres med 17 skills i Claude-Code-stilens `/SKILL.md`-format: 9 kaldbare action-skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) og 8 referenceskills, som agenten indlæser efter behov (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Referenceskillene indeholder datatabeller genereret fra kildekoden, så de aldrig driver ud af kurs. [`skills`](https://npmjs.com/package/skills)-CLI'en fra vercel-labs installerer dem automatisk i den kaldende agents native skill-mappe på tværs af 50+ agenter (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf og flere):
@@ -836,7 +628,7 @@ agentmemory-indgangen er den **samme MCP-serverblok** på alle hosts, der bruger
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` installerer MCP og capture-hooks i den delte tilpasningsmappe. Se [Antigravity-opsætning og begrænsninger](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` bruger samme MCP- og hook-konfiguration som de nuværende IDE-versioner. Eksisterende installationer bør opdateres med `--force`; se [opgraderingsnoterne](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` skriver konfigurationen på brugerniveau. Workspace-overrides placeres i `.kiro/settings/mcp.json` ved siden af din kode. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` skriver standard-`mcpServers`-blokken. Warp opdager også automatisk skills fra `.claude/skills/`; når Claude Code-pluginet er installeret, vises de 8 agentmemory-skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) nativt i Warps slash-kommando-palette. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` skriver standard-`mcpServers`-blokken. Warp opdager også automatisk skills fra `.claude/skills/`; når Claude Code-pluginet er installeret, vises de 9 agentmemory-skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) nativt i Warps slash-kommando-palette. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` skriver standard-`mcpServers`-blokken. Brugere af VS Code-extensionen: indsæt samme blok via Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (foretrukket) eller `config.json` (legacy) | `agentmemory connect continue` opretter en manglende `config.yaml` eller ændrer eksisterende `config.json`. Hvis YAML-filen findes, vises blokken til `mcpServers:`, så kommentarer og ankre bevares. Continue bruger et array, ikke et objekt. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` skriver under `context_servers` (Zeds nøgle, IKKE `mcpServers`). Eksterne MCP-servere kan i stedet forbindes via `{"url": "..."}`. |
@@ -933,7 +725,6 @@ Hver AI-kodeagent leveres med indbygget hukommelse: Claude Code har `MEMORY.md`,
|---|---|---|
| Vedligeholdelse | Manuelt vedligeholdte instruktioner | Indekseret sessionshistorik og hukommelse |
| Søgning | Indlæser alt i konteksten | BM25 + vektor + graf (kun top-K) |
-| Tokenomkostning | 22K+ ved 240 observationer | ~1,900 tokens (92% mindre) |
| Tværs af agenter | Filer pr. agent | MCP + REST (enhver agent) |
| Koordinering | Ingen | Leases, signaler, actions, routines |
| Observerbarhed | Læs filer manuelt | Realtids-viewer på :3113 |
@@ -946,7 +737,7 @@ Hver AI-kodeagent leveres med indbygget hukommelse: Claude Code har `MEMORY.md`,
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1031,7 +822,7 @@ Fusioneret med Reciprocal Rank Fusion (RRF, k=60) og session-diversificeret (mak
Når et vektorindeks er udfyldt, bruger `mem::search` (bag `memory_recall`) den hybride BM25 + vektor-ranker. Uden embeddings bruger den BM25. `smart-search` kan desuden fusionere strukturelle graf-match, når der findes graf-data, også i nøgleløs tilstand. Lesson-recall køres på et dedikeret in-memory BM25-indeks i stedet for at scanne hele korpuset pr. forespørgsel. Overskrevne hukommelsesversioner udelukkes fra alle recall-veje; versionskæden bevarer deres historik.
-Vektorer overlever et crash eller force-kill. Vektorindekset gemmes i buckets højst hver `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minutter). Hver vektor, der tilføjes eller fjernes imellemtiden, skrives også straks til en lille ventende log i state-store'en, og den næste opstart afspiller den uden at kalde embedding-udbyderen. Hver gennemført save tømmer loggen. Dokumenter, der stadig ikke har en vektor efter afspilningen, genindlejres i baggrunden i batches af `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500), indtil ingen er tilbage, og en backfill, der stoppes, fortsætter ved næste opstart. `/agentmemory/status` og viewer viser størrelsen af den ventende log og backfill-tilstanden. Nøgleløse installationer skriver ingenting.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokeniserer græsk, kyrillisk, hebraisk, arabisk og accentueret latin ud af boksen. For kinesiske/japanske/koreanske hukommelser skal du installere de valgfrie segmenterere (`npm install @node-rs/jieba tiny-segmenter`) for at splitte CJK-forløb i ord-niveau-tokens; uden dem falder agentmemory blidt tilbage til hel-forløb-tokenisering og printer et engangshint på stderr.
@@ -1047,9 +838,9 @@ Den normale npm-installation inkluderer den valgfrie `@huggingface/transformers`
| Udbyder | Model | Pris | Noter |
|---|---|---|---|
-| **Lokal (anbefalet tilvalg)** | `all-MiniLM-L6-v2` | Gratis | På enheden efter første modeldownload, +8pp recall over kun BM25 |
+| **Lokal (anbefalet tilvalg)** | `all-MiniLM-L6-v2` | Gratis | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Gratis niveau | 100+ sprog, 768/1536/3072 dimensioner (MRL), 2048-token input. Afløser `text-embedding-004` ([udfaset, lukket 14. januar 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Højeste kvalitet |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Betalt | Optimeret til kode |
| Cohere | `embed-english-v3.0` | Gratis prøveperiode | Generelt anvendelig |
| OpenRouter | Enhver model | Varierer | Multi-model-proxy |
diff --git a/READMEs/README.de-DE.md b/READMEs/README.de-DE.md
index 1641333254..311e1459ee 100644
--- a/READMEs/README.de-DE.md
+++ b/READMEs/README.de-DE.md
@@ -66,8 +66,6 @@
-
-> Embedding-Modell: `all-MiniLM-L6-v2` (lokal, kostenlos, kein API-Schlüssel). Vollständige Berichte: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Konkurrenzvergleich: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md), der agentmemory vs mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo abdeckt.
-
-**Lokal reproduzieren:** [`eval/README.md`](../eval/README.md), ein Adapter-pluggable Harness für LongMemEval `_s` (öffentlich, 500 Fragen) + `coding-agent-life-v1` (interner 15-Session-Korpus). Adapter für Grep / Vector / agentmemory werden direkt verglichen, NDJSON-Ausgabe, veröffentlichte Scorecards landen in [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Funktioniert kombiniert mit [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) und [Graphify](https://github.com/safishamsi/graphify).** Code-Graph-Indizierung, mehragentige Build-Pipelines und breitere Knowledge Graphs über Docs / PDFs / Bilder / Videos. agentmemory merkt sich die Arbeit; diese drei Projekte beleuchten den Rest der Kontextschicht. Rezepte + Frage-Routing-Tabelle: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Eingebaut (CLAUDE.md)
-
-
-
Typ
-
Memory-Engine + MCP-Server
-
Memory-Layer-API
-
Komplette Agenten-Runtime
-
Persönliche KI
-
Memory-API + App
-
Team-Memory-Hub (LLM-Proxy)
-
Vector-Memory (OSS)
-
Memory-Engine (Oracle DB)
-
Memory-System
-
Statische Datei
-
-
-
Retrieval R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/V
-
Selbst berichtet
-
PersonaMem 76% (selbst berichtet)
-
~96.6% (selbst berichtet)
-
94.4% (selbst berichtet)
-
N/V
-
N/V (grep)
-
-
-
Auto-Erfassung
-
12 Hooks (null manueller Aufwand)
-
Manuelle add()-Aufrufe
-
Agent bearbeitet sich selbst
-
Manuell
-
API-seitige Extraktion
-
Proxy-Interception (Base-URL-Tausch)
-
Manuell
-
API-Extraktion
-
Manuell
-
Manuelle Bearbeitung
-
-
-
Suche
-
BM25 + Vector + Graph (RRF-Fusion)
-
Vector + Graph
-
Vector (Archival)
-
Semantisch
-
Vector + RAG
-
4 Asset-Typen (Chat / Skill / Wiki / CodeGraph)
-
Nur Vector
-
Vector + semantisch
-
Decay-gewichtet
-
Lädt alles in den Kontext
-
-
-
Multi-Agent
-
MCP + REST + Leases + Signals
-
API (keine Koordination)
-
Nur innerhalb der Letta-Runtime
-
Nein
-
Nein
-
Team-Rollen + geteilte Assets
-
Nein
-
Nur Scoped
-
Multi-Agent geteilt
-
Dateien pro Agent
-
-
-
Framework-Lock-in
-
Keiner (jeder MCP-Client)
-
Keiner
-
Hoch (Letta erforderlich)
-
Standalone
-
Keiner
-
Proxy sitzt vor jedem Modellaufruf
-
Keiner
-
Oracle Database
-
Keiner
-
Format pro Agent
-
-
-
Externe Abhängigkeiten
-
Kein separater Datenbankdienst erforderlich
-
Qdrant / pgvector
-
Postgres + Vector-DB
-
Mehrere
-
Managed Cloud
-
Docker-Stack (Core + Hub + Proxy)
-
Vector-Store
-
Oracle AI Database
-
Keine
-
Keine
-
-
-
Memory-Lifecycle
-
4-stufige Konsolidierung + Decay + Auto-Forget
-
Passive Extraktion
-
Vom Agenten verwaltet
-
Manuell
-
Auto-Forget
-
Manuelles Review; Auto-Routing in Arbeit
-
Keiner
-
Nicht angegeben
-
Decay + Konsolidierung
-
Manuelles Pruning
-
-
-
Token-Effizienz
-
~1,900 Tokens/Session ($10/Jahr)
-
Je nach Integration unterschiedlich
-
Core Memory im Kontext
-
Variiert
-
Cloud-Preise
-
Nicht angegeben
-
Kein Token-Budget
-
LLM-gestützt (variiert)
-
Variiert
-
22K+ Tokens bei 240 Beobachtungen
-
-
-
Echtzeit-Viewer
-
Ja (Port 3113)
-
Cloud-Dashboard
-
Cloud-Dashboard
-
Web-UI
-
Cloud-Dashboard
-
Hub-Web-UI
-
Nein
-
Nein
-
Nein
-
Nein
-
-
-
Self-hosted
-
Ja (Standard)
-
Optional
-
Optional
-
Ja
-
Nein (nur Cloud)
-
Ja (Docker)
-
Ja
-
Ja (Oracle DB)
-
Ja
-
Ja
-
-
-
-Benchmark-Hinweis: Nur agentmemorys R@5 ist unser eigenes gemessenes Ergebnis (LongMemEval-S, reproduzierbar aus benchmark/COMPARISON.md). Die Zahlen von mem0 und Letta sind deren veröffentlichte LoCoMo-Werte (ein anderer Datensatz); die Zahlen von MemPalace, supermemory, TencentDB (PersonaMem) und oracleagentmemory sind selbst berichtete Herstellerangaben, die wir nicht unabhängig reproduziert haben (der Lauf von oracleagentmemory verwendete GPT-5.5 gegen eine Oracle AI Database). Nebeneinander nur zur groben Einordnung gezeigt, kein direkter Vergleich auf identischen Daten. Star-Zahlen sind ungefähr und driften über die Zeit.
-
-**Neuere Einsteiger**, die man kennen sollte, ausführlich verglichen in [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| System | ⭐ | Ausrichtung |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporaler Knowledge Graph; stärkste veröffentlichte Temporal-Query-Ergebnisse (LongMemEval 63.8%), aber der Graph wird asynchron aufgebaut, sodass frische Fakten hinterherhinken können |
-| Cognee | 30K | Dokument-zu-Knowledge-Graph-Ingestion, nur Python, gebaut für strukturierte Entitäten-Extraktion statt Session-Erfassung |
+Aktueller Vergleich von Produkten und Bereitstellungsoptionen mit Quellen:
-AgentMemory verbindet die Erfassung durch Coding-Agenten mit durchsuchbarer Projekthistorie und einem lokalen Viewer. Der Betrieb ohne API-Schlüssel ist vom ersten Start an möglich.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -777,7 +569,7 @@ Vollständiger Leitfaden: [`integrations/hermes/`](../integrations/hermes/)
Starten Sie den Memory-Server: `npx -y @agentmemory/agentmemory@latest`
-#### Native Skills via `npx skills add` (50+ Agenten)
+#### Native Skills via `npx skills add`
agentmemory liefert 17 Skills im Claude-Code-artigen `/SKILL.md`-Format: 9 aufrufbare Action-Skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) und 8 Referenz-Skills, die der Agent bei Bedarf lädt (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Die Referenz-Skills tragen aus dem Quellcode generierte Datentabellen, sodass sie nie driften. Die [`skills`](https://npmjs.com/package/skills)-CLI von vercel-labs installiert sie automatisch in das native Skill-Verzeichnis des aufrufenden Agenten, über 50+ Agenten hinweg (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf und mehr):
@@ -834,7 +626,7 @@ Der agentmemory-Eintrag ist der **gleiche MCP-Server-Block** für jeden Host, de
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` installiert MCP und Capture-Hooks im gemeinsam genutzten Customization-Verzeichnis. Siehe [Antigravity-Setup und -Grenzen](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` verwendet dieselbe MCP- und Hook-Konfiguration wie aktuelle IDE-Versionen. Bestehende Installationen sollten mit `--force` aktualisieren; siehe die [Upgrade-Hinweise](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` schreibt die Konfig auf Benutzerebene. Workspace-Overrides liegen in `.kiro/settings/mcp.json` neben Ihrem Code. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` schreibt den standardmäßigen `mcpServers`-Block. Warp entdeckt außerdem Skills aus `.claude/skills/` automatisch; sobald das Claude-Code-Plugin installiert ist, erscheinen die 8 agentmemory-Skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) nativ in Warps Slash-Command-Palette. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` schreibt den standardmäßigen `mcpServers`-Block. Warp entdeckt außerdem Skills aus `.claude/skills/` automatisch; sobald das Claude-Code-Plugin installiert ist, erscheinen die 9 agentmemory-Skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) nativ in Warps Slash-Command-Palette. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` schreibt den standardmäßigen `mcpServers`-Block. Nutzer der VS-Code-Extension: Fügen Sie denselben Block über Cline Settings → MCP Servers → Edit JSON ein. |
| **Continue.dev** | `~/.continue/config.yaml` (bevorzugt) oder `config.json` (Legacy) | `agentmemory connect continue` erstellt eine fehlende `config.yaml` oder aktualisiert eine vorhandene `config.json`. Bei bestehender YAML-Datei zeigt der Adapter den Block für `mcpServers:` zum Einfügen an; Kommentare und Anker bleiben erhalten. Continue verwendet dort ein Array, kein Objekt. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` schreibt unter `context_servers` (Zeds Schlüssel, NICHT `mcpServers`). Remote-MCP-Server können stattdessen via `{"url": "..."}` verdrahtet werden. |
@@ -931,7 +723,6 @@ Jeder KI-Coding-Agent kommt mit eingebautem Memory: Claude Code hat `MEMORY.md`,
|---|---|---|
| Pflege | Manuell gepflegte Anweisungen | Indizierte Session-Historie und Erinnerungen |
| Suche | Lädt alles in den Kontext | BM25 + Vector + Graph (nur Top-K) |
-| Token-Kosten | 22K+ bei 240 Beobachtungen | ~1,900 Tokens (92 % weniger) |
| Agentenübergreifend | Dateien pro Agent | MCP + REST (jeder Agent) |
| Koordination | Keine | Leases, Signale, Actions, Routinen |
| Observability | Dateien manuell lesen | Echtzeit-Viewer auf :3113 |
@@ -944,7 +735,7 @@ Jeder KI-Coding-Agent kommt mit eingebautem Memory: Claude Code hat `MEMORY.md`,
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1029,7 +820,7 @@ Verschmolzen mit Reciprocal Rank Fusion (RRF, k=60) und session-diversifiziert (
Wenn ein Vector-Index befüllt ist, verwendet `mem::search` (hinter `memory_recall`) den hybriden BM25-+-Vector-Ranker. Ohne Embeddings verwendet es BM25. `smart-search` kann zusätzlich strukturelle Graph-Treffer einbeziehen, wenn Graph-Daten existieren, auch im schlüssellosen Modus. Lesson-Recall läuft auf einem dedizierten In-Memory-BM25-Index, statt bei jeder Anfrage den ganzen Korpus zu scannen. Überholte Memory-Versionen sind von jedem Recall-Pfad ausgeschlossen; die Versionskette bewahrt ihre Historie.
-Vectors überleben einen Absturz oder Force-Kill. Der Vector-Index wird in Buckets gespeichert, höchstens alle `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 Minuten). Jeder dazwischen hinzugefügte oder entfernte Vector wird außerdem sofort in ein kleines Pending-Log im State-Store geschrieben, und der nächste Start spielt es ab, ohne den Embedding-Provider aufzurufen. Jedes erfolgreiche Speichern leert das Log. Dokumente, die nach dem Replay noch keinen Vector haben, werden im Hintergrund in Batches von `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) neu eingebettet, bis keine mehr übrig sind, und ein gestopptes Backfill wird beim nächsten Start fortgesetzt. `/agentmemory/status` und der Viewer zeigen die Größe des Pending-Logs und den Backfill-Status. Schlüssellose Installationen schreiben nichts.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokenisiert Griechisch, Kyrillisch, Hebräisch, Arabisch und akzentuiertes Latein standardmäßig. Für Erinnerungen in Chinesisch / Japanisch / Koreanisch installieren Sie die optionalen Segmentierer (`npm install @node-rs/jieba tiny-segmenter`), um CJK-Folgen in Worttokens aufzuteilen; ohne sie fällt agentmemory weich auf eine Tokenisierung als gesamte Folge zurück und gibt einmalig einen Hinweis auf stderr aus.
@@ -1045,9 +836,9 @@ Die normale npm-Installation enthält die optionale `@huggingface/transformers`-
| Provider | Modell | Kosten | Hinweise |
|---|---|---|---|
-| **Lokal (empfohlenes Opt-in)** | `all-MiniLM-L6-v2` | Kostenlos | On-device nach dem ersten Modell-Download, +8pp Recall gegenüber Nur-BM25 |
+| **Lokal (empfohlenes Opt-in)** | `all-MiniLM-L6-v2` | Kostenlos | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Free Tier | 100+ Sprachen, 768/1536/3072 Dims (MRL), 2048-Token-Eingabe. Ersetzt `text-embedding-004` ([deprecated, Abschaltung 14. Jan. 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Höchste Qualität |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Kostenpflichtig | Auf Code optimiert |
| Cohere | `embed-english-v3.0` | Testzugang | Allzweck |
| OpenRouter | Beliebiges Modell | Variabel | Multi-Modell-Proxy |
diff --git a/READMEs/README.el-GR.md b/READMEs/README.el-GR.md
index fa2254a851..a021d9cfa9 100644
--- a/READMEs/README.el-GR.md
+++ b/READMEs/README.el-GR.md
@@ -66,8 +66,6 @@
-
-> Μοντέλο embedding: `all-MiniLM-L6-v2` (τοπικό, δωρεάν, χωρίς κλειδί API). Πλήρεις αναφορές: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Σύγκριση με ανταγωνιστές: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md), που καλύπτει το agentmemory έναντι των mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Αναπαραγωγή τοπικά:** [`eval/README.md`](../eval/README.md), ένα harness με συνδέσιμους adapters για το LongMemEval `_s` (δημόσιο, 500 ερωτήσεις) + `coding-agent-life-v1` (εσωτερικό corpus 15 συνεδριών). Οι adapters grep / vector / agentmemory βαθμολογούνται παράλληλα, με έξοδο NDJSON, και τα δημοσιευμένα scorecards καταλήγουν στο [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Συνεργάζεται με τα [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything), και [Graphify](https://github.com/safishamsi/graphify).** Ευρετηρίαση code-graph, pipelines κατασκευής πολλών agents, και ευρύτεροι γράφοι γνώσης σε docs / PDF / εικόνες / βίντεο. Το agentmemory θυμάται τη δουλειά· αυτά τα τρία projects αναδεικνύουν την υπόλοιπη context layer. Συνταγές + πίνακας δρομολόγησης ερωτημάτων: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Ενσωματωμένο (CLAUDE.md)
-
-
-
Τύπος
-
Μηχανή μνήμης + διακομιστής MCP
-
API επιπέδου μνήμης
-
Πλήρες runtime agent
-
Προσωπικό AI
-
API μνήμης + εφαρμογή
-
Κόμβος μνήμης ομάδας (proxy LLM)
-
Vector μνήμη (OSS)
-
Μηχανή μνήμης (Oracle DB)
-
Σύστημα μνήμης
-
Στατικό αρχείο
-
-
-
Ανάκληση R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Αυτοαναφερόμενο
-
PersonaMem 76% (αυτοαναφερόμενο)
-
~96.6% (αυτοαναφερόμενο)
-
94.4% (αυτοαναφερόμενο)
-
N/A
-
N/A (grep)
-
-
-
Αυτόματη καταγραφή
-
12 hooks (καμία χειροκίνητη προσπάθεια)
-
Χειροκίνητες κλήσεις add()
-
Αυτο-επεξεργασίες από τον agent
-
Χειροκίνητο
-
Εξαγωγή από την πλευρά του API
-
Παρεμβολή proxy (αλλαγή base-URL)
-
Χειροκίνητο
-
Εξαγωγή API
-
Χειροκίνητο
-
Χειροκίνητη επεξεργασία
-
-
-
Αναζήτηση
-
BM25 + Vector + Graph (συγχώνευση RRF)
-
Vector + Graph
-
Vector (αρχειοθέτηση)
-
Σημασιολογική
-
Vector + RAG
-
4 τύποι assets (Chat / Skill / Wiki / CodeGraph)
-
Μόνο Vector
-
Vector + σημασιολογική
-
Σταθμισμένη βάσει decay
-
Φορτώνει τα πάντα στο context
-
-
-
Πολλαπλών agents
-
MCP + REST + leases + σήματα
-
API (χωρίς συντονισμό)
-
Μόνο εντός του runtime Letta
-
Όχι
-
Όχι
-
Ρόλοι ομάδας + κοινά assets
-
Όχι
-
Μόνο εντός scope
-
Κοινό μεταξύ πολλών agents
-
Αρχεία ανά agent
-
-
-
Εξάρτηση από framework
-
Καμία (οποιοδήποτε MCP client)
-
Καμία
-
Υψηλή (απαιτεί χρήση Letta)
-
Αυτόνομο
-
Καμία
-
Το proxy μεσολαβεί σε κάθε κλήση μοντέλου
-
Καμία
-
Oracle Database
-
Καμία
-
Μορφή ανά agent
-
-
-
Εξωτερικές εξαρτήσεις
-
Δεν απαιτείται ξεχωριστή υπηρεσία βάσης δεδομένων
-
Qdrant / pgvector
-
Postgres + vector DB
-
Πολλαπλές
-
Διαχειριζόμενο cloud
-
Docker stack (Core + Hub + Proxy)
-
Αποθήκη vector
-
Oracle AI Database
-
Καμία
-
Καμία
-
-
-
Κύκλος ζωής μνήμης
-
Ενοποίηση 4 επιπέδων + decay + αυτόματη διαγραφή
-
Παθητική εξαγωγή
-
Διαχειρίζεται από τον agent
-
Χειροκίνητο
-
Αυτόματη διαγραφή
-
Χειροκίνητη επισκόπηση· αυτόματη δρομολόγηση σε εξέλιξη
-
Καμία
-
Δεν αναφέρεται
-
Decay + ενοποίηση
-
Χειροκίνητο κλάδεμα
-
-
-
Απόδοση tokens
-
~1,900 tokens/συνεδρία ($10/έτος)
-
Διαφέρει ανά ολοκλήρωση
-
Βασική μνήμη εντός context
-
Διαφέρει
-
Τιμολόγηση cloud
-
Δεν αναφέρεται
-
Χωρίς προϋπολογισμό tokens
-
Βασισμένο σε LLM (διαφέρει)
-
Διαφέρει
-
22K+ tokens σε 240 παρατηρήσεις
-
-
-
Viewer σε πραγματικό χρόνο
-
Ναι (θύρα 3113)
-
Dashboard cloud
-
Dashboard cloud
-
Web UI
-
Dashboard cloud
-
Hub web UI
-
Όχι
-
Όχι
-
Όχι
-
Όχι
-
-
-
Αυτοφιλοξενούμενο (self-hosted)
-
Ναι (προεπιλογή)
-
Προαιρετικό
-
Προαιρετικό
-
Ναι
-
Όχι (μόνο cloud)
-
Ναι (Docker)
-
Ναι
-
Ναι (Oracle DB)
-
Ναι
-
Ναι
-
-
-
-Σημείωση benchmark: μόνο το R@5 του agentmemory είναι δικό μας μετρημένο αποτέλεσμα (LongMemEval-S, αναπαραγώγιμο από το benchmark/COMPARISON.md). Τα στοιχεία των mem0 και Letta είναι οι δημοσιευμένες τιμές τους στο LoCoMo (διαφορετικό dataset)· τα στοιχεία των MemPalace, supermemory, TencentDB (PersonaMem), και oracleagentmemory είναι αυτοαναφερόμενοι ισχυρισμοί προμηθευτών που δεν έχουμε αναπαράγει ανεξάρτητα (η εκτέλεση του oracleagentmemory χρησιμοποίησε GPT-5.5 έναντι μιας Oracle AI Database). Παρουσιάζονται δίπλα-δίπλα μόνο για κατά προσέγγιση σύγκριση, όχι ως απευθείας αντιπαράθεση σε ταυτόσημα δεδομένα. Οι μετρήσεις αστεριών είναι κατά προσέγγιση και μεταβάλλονται με τον χρόνο.
-
-**Νεότεροι παίκτες** που αξίζει να γνωρίζετε, συγκρίνονται αναλυτικά στο [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Σύστημα | ⭐ | Προσέγγιση |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Χρονικός γράφος γνώσης· τα ισχυρότερα δημοσιευμένα αποτελέσματα σε χρονικά ερωτήματα (LongMemEval 63.8%), αλλά ο γράφος χτίζεται ασύγχρονα, οπότε τα νέα δεδομένα μπορεί να καθυστερούν |
-| Cognee | 30K | Εισαγωγή εγγράφων σε γράφο γνώσης, μόνο σε Python, σχεδιασμένο για δομημένη εξαγωγή οντοτήτων παρά για καταγραφή συνεδριών |
+Ενημερωμένη σύγκριση προϊόντων και επιλογών εγκατάστασης με πηγές:
-Το AgentMemory συνδυάζει καταγραφή συνεδριών, αναζητήσιμο ιστορικό έργου και τοπικό viewer, με λειτουργία χωρίς κλειδί από την πρώτη εκκίνηση.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -786,7 +578,7 @@ Verify with `curl http://localhost:3111/agentmemory/health`. Open http://localho
Ξεκινήστε τον διακομιστή μνήμης: `npx -y @agentmemory/agentmemory@latest`
-#### Εγγενή skills μέσω `npx skills add` (50+ agents)
+#### Εγγενή skills μέσω `npx skills add`
Το agentmemory διανέμει 17 skills στη μορφή `/SKILL.md` τύπου Claude Code: 9 επικλητά skills ενέργειας (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) και 8 skills αναφοράς που ο agent φορτώνει όταν τα χρειάζεται (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Τα skills αναφοράς περιέχουν πίνακες δεδομένων που δημιουργούνται από την πηγή, οπότε δεν αποκλίνουν ποτέ. Το CLI [`skills`](https://npmjs.com/package/skills) από την vercel-labs τα εγκαθιστά αυτόματα στον εγγενή κατάλογο skills του agent που το καλεί, σε 50+ agents (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, και άλλα):
@@ -844,7 +636,7 @@ npx skills add rohitg00/agentmemory -y -a '*' # install to every installed age
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | Το `agentmemory connect antigravity --with-hooks` εγκαθιστά το MCP και τα hooks καταγραφής στον κοινόχρηστο κατάλογο προσαρμογής. Δείτε [ρύθμιση και περιορισμοί Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | Το `agentmemory connect antigravity-cli --with-hooks` χρησιμοποιεί την ίδια ρύθμιση MCP και hooks με τις τρέχουσες εκδόσεις IDE. Οι υπάρχουσες εγκαταστάσεις πρέπει να ανανεωθούν με `--force`· δείτε τις [σημειώσεις αναβάθμισης](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | Το `agentmemory connect kiro` γράφει τις ρυθμίσεις σε επίπεδο χρήστη. Οι παρακάμψεις workspace μπαίνουν στο `.kiro/settings/mcp.json` δίπλα στον κώδικά σας. |
-| **Warp** | `~/.warp/.mcp.json` | Το `agentmemory connect warp` γράφει το τυπικό block `mcpServers`. Το Warp επίσης ανακαλύπτει αυτόματα skills από το `.claude/skills/`· μόλις εγκατασταθεί το plugin του Claude Code, τα 8 skills του agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) εμφανίζονται εγγενώς στην παλέτα εντολών slash του Warp. |
+| **Warp** | `~/.warp/.mcp.json` | Το `agentmemory connect warp` γράφει το τυπικό block `mcpServers`. Το Warp επίσης ανακαλύπτει αυτόματα skills από το `.claude/skills/`· μόλις εγκατασταθεί το plugin του Claude Code, τα 9 skills του agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) εμφανίζονται εγγενώς στην παλέτα εντολών slash του Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | Το `agentmemory connect cline` γράφει το τυπικό block `mcpServers`. Χρήστες της extension VS Code: επικολλήστε το ίδιο block μέσω Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (προτιμώμενο) ή `config.json` (παλαιό) | Το `agentmemory connect continue` δημιουργεί `config.yaml` αν λείπουν οι ρυθμίσεις ή αλλάζει υπάρχον `config.json`. Για υπάρχον YAML εμφανίζει το block προς προσθήκη στο `mcpServers:`, διατηρώντας σχόλια και anchors. Το Continue χρησιμοποιεί πίνακα, όχι αντικείμενο. |
| **Zed** | `~/.config/zed/settings.json` | Το `agentmemory connect zed` γράφει κάτω από το `context_servers` (το κλειδί του Zed, ΟΧΙ `mcpServers`). Απομακρυσμένοι διακομιστές MCP μπορούν να συνδεθούν εναλλακτικά μέσω `{"url": "..."}`. |
@@ -927,7 +719,6 @@ npx -y @agentmemory/mcp
|---|---|---|
| Συντήρηση | Χειροκίνητα διαχειριζόμενες οδηγίες | Ευρετηριασμένο ιστορικό συνεδριών και μνήμες |
| Αναζήτηση | Φορτώνει τα πάντα στο context | BM25 + vector + graph (μόνο top-K) |
-| Κόστος tokens | 22K+ σε 240 παρατηρήσεις | ~1,900 tokens (92% λιγότερο) |
| Μεταξύ agents | Αρχεία ανά agent | MCP + REST (οποιοσδήποτε agent) |
| Συντονισμός | Κανένας | Leases, signals, actions, routines |
| Παρατηρησιμότητα | Χειροκίνητη ανάγνωση αρχείων | Viewer σε πραγματικό χρόνο στη :3113 |
@@ -940,7 +731,7 @@ npx -y @agentmemory/mcp
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1025,7 +816,7 @@ SessionStart hook fires
Όταν ένα ευρετήριο vector είναι γεμάτο, το `mem::search` (πίσω από το `memory_recall`) χρησιμοποιεί τον υβριδικό ranker BM25 + vector. Χωρίς embeddings χρησιμοποιεί BM25. Το `smart-search` μπορεί επιπλέον να συγχωνεύσει δομικές αντιστοιχίσεις γράφου όταν υπάρχουν δεδομένα γράφου, ακόμα και σε λειτουργία χωρίς κλειδί. Η ανάκληση μαθημάτων (lesson recall) τρέχει σε ένα αποκλειστικό in-memory ευρετήριο BM25 αντί να σκανάρει όλο το corpus σε κάθε ερώτημα. Οι αντικατεστημένες εκδόσεις μνήμης αποκλείονται από κάθε διαδρομή ανάκλησης· η αλυσίδα εκδόσεων διατηρεί το ιστορικό τους.
-Τα vectors επιβιώνουν από ένα crash ή force-kill. Το ευρετήριο vector αποθηκεύεται σε buckets το πολύ κάθε `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 λεπτά). Κάθε vector που προστίθεται ή αφαιρείται στο ενδιάμεσο γράφεται επίσης αμέσως σε ένα μικρό εκκρεμές log στο state store, και η επόμενη εκκίνηση το αναπαράγει χωρίς να καλέσει τον πάροχο embedding. Κάθε επιτυχημένη αποθήκευση αδειάζει το log. Τα έγγραφα που εξακολουθούν να μην έχουν vector μετά την αναπαραγωγή επανα-embeddάρονται στο παρασκήνιο σε batches των `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) μέχρι να μην απομείνει κανένα, και ένα backfill που διακόπτεται συνεχίζει στην επόμενη εκκίνηση. Το `/agentmemory/status` και το viewer δείχνουν το μέγεθος του εκκρεμούς log και την κατάσταση του backfill. Οι εγκαταστάσεις χωρίς κλειδί δεν γράφουν τίποτα.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
Το BM25 κάνει tokenize Ελληνικά, Κυριλλικά, Εβραϊκά, Αραβικά, και τονισμένα Λατινικά απευθείας. Για μνήμες στα Κινεζικά / Ιαπωνικά / Κορεατικά, εγκαταστήστε τους προαιρετικούς segmenters (`npm install @node-rs/jieba tiny-segmenter`) για να χωρίσετε τις ακολουθίες CJK σε tokens επιπέδου λέξης· χωρίς αυτούς, το agentmemory υποχωρεί ομαλά σε tokenization ολόκληρης της ακολουθίας και εκτυπώνει μία υπόδειξη, μία φορά, στο stderr.
@@ -1041,9 +832,9 @@ EMBEDDING_PROVIDER=local
| Πάροχος | Μοντέλο | Κόστος | Σημειώσεις |
|---|---|---|---|
-| **Τοπικό (προτεινόμενη ενεργοποίηση)** | `all-MiniLM-L6-v2` | Δωρεάν | Στη συσκευή μετά την πρώτη λήψη μοντέλου, +8pp ανάκληση σε σχέση με μόνο BM25 |
+| **Τοπικό (προτεινόμενη ενεργοποίηση)** | `all-MiniLM-L6-v2` | Δωρεάν | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Δωρεάν βαθμίδα | 100+ γλώσσες, διαστάσεις 768/1536/3072 (MRL), είσοδος 2048 tokens. Αντικαθιστά το `text-embedding-004` ([καταργήθηκε, διακοπή στις 14 Ιαν 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Υψηλότερη ποιότητα |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Επί πληρωμή | Βελτιστοποιημένο για κώδικα |
| Cohere | `embed-english-v3.0` | Δωρεάν δοκιμή | Γενικής χρήσης |
| OpenRouter | Οποιοδήποτε μοντέλο | Διαφέρει | Proxy πολλών μοντέλων |
diff --git a/READMEs/README.es-ES.md b/READMEs/README.es-ES.md
index a18cec0e78..a66ef720a8 100644
--- a/READMEs/README.es-ES.md
+++ b/READMEs/README.es-ES.md
@@ -66,8 +66,6 @@
-
-
@@ -331,223 +329,17 @@ Notas de la última release: [CHANGELOG.md](../CHANGELOG.md).
-
-
-
-
-### Precisión de recuperación
-
-**coding-agent-life-v1** (corpus propio, reproducible en sandbox)
-
-| Adaptador | P@5 | R@5 | Tasa de aciertos top-5 | Latencia p50 |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| baseline grep | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-Tasa de aciertos top-5 del 100% en el **techo matemático de P@5** para este corpus (0.240, ver scorecard). Hybrid recupera todas las sesiones gold; grep falla 1 de 2 gold en la consulta temporal multi-sesión. La mejora es **recall + temporal**, no precisión agregada. Este benchmark es pequeño y escaso en gold; el LongMemEval-S más grande de abajo diferencia mejor. Desglose completo por tipo + nota de corrección: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 preguntas)
+Resultados actuales de recuperación, definiciones de métricas y pasos para reproducirlos:
-| Sistema | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| Fallback solo BM25 | 86.2% | 94.6% | 71.5% |
-
-
-
-> Modelo de embedding: `all-MiniLM-L6-v2` (local, gratis, sin API key). Informes completos: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Comparativa con competidores: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md), que cubre agentmemory frente a mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Reprodúcelo en local:** [`eval/README.md`](../eval/README.md), un harness con adaptadores intercambiables para LongMemEval `_s` (500 preguntas públicas) + `coding-agent-life-v1` (corpus propio de 15 sesiones). Los adaptadores grep / vector / agentmemory puntúan uno junto al otro, salida NDJSON, los scorecards publicados aterrizan en [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Funciona muy bien con [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) y [Graphify](https://github.com/safishamsi/graphify).** Indexado de grafos de código, pipelines de build multiagente y grafos de conocimiento más amplios sobre documentos / PDFs / imágenes / vídeos. agentmemory recuerda el trabajo; esos tres proyectos iluminan el resto de la capa de contexto. Recetas y tabla de enrutamiento por pregunta: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Integrada (CLAUDE.md)
-
-
-
Tipo
-
Motor de memoria + servidor MCP
-
API de capa de memoria
-
Runtime de agente completo
-
IA personal
-
API de memoria + app
-
Hub de memoria de equipo (proxy LLM)
-
Memoria vectorial (OSS)
-
Motor de memoria (Oracle DB)
-
Sistema de memoria
-
Fichero estático
-
-
-
R@5 de recuperación
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Autoinformado
-
PersonaMem 76% (autoinformado)
-
~96.6% (autoinformado)
-
94.4% (autoinformado)
-
N/A
-
N/A (grep)
-
-
-
Captura automática
-
12 hooks (esfuerzo manual cero)
-
Llamadas manuales a add()
-
Autoedición por el agente
-
Manual
-
Extracción del lado de la API
-
Intercepción por proxy (cambio de base-URL)
-
Manual
-
Extracción por API
-
Manual
-
Edición manual
-
-
-
Búsqueda
-
BM25 + Vector + Graph (fusión RRF)
-
Vector + Graph
-
Vector (archival)
-
Semántica
-
Vector + RAG
-
4 tipos de asset (Chat / Skill / Wiki / CodeGraph)
-
Solo vector
-
Vector + semántica
-
Ponderada por decaimiento
-
Carga todo en el contexto
-
-
-
Multiagente
-
MCP + REST + leases + signals
-
API (sin coordinación)
-
Solo dentro del runtime de Letta
-
No
-
No
-
Roles de equipo + assets compartidos
-
No
-
Solo con scope
-
Compartido multiagente
-
Ficheros por agente
-
-
-
Lock-in de framework
-
Ninguno (cualquier cliente MCP)
-
Ninguno
-
Alto (debes usar Letta)
-
Standalone
-
Ninguno
-
Un proxy antepone cada llamada al modelo
-
Ninguno
-
Oracle Database
-
Ninguno
-
Formato por agente
-
-
-
Dependencias externas
-
No necesita un servicio de base de datos separado
-
Qdrant / pgvector
-
Postgres + BD vectorial
-
Varias
-
Nube gestionada
-
Stack Docker (Core + Hub + Proxy)
-
Vector store
-
Oracle AI Database
-
Ninguna
-
Ninguna
-
-
-
Ciclo de vida de la memoria
-
Consolidación de 4 niveles + decaimiento + auto-olvido
-
Extracción pasiva
-
Gestionado por el agente
-
Manual
-
Auto-olvido
-
Revisión manual; auto-enrutado en progreso
-
Ninguno
-
No especificado
-
Decaimiento + consolidación
-
Poda manual
-
-
-
Eficiencia de tokens
-
~1,900 tokens/sesión ($10/año)
-
Varía según la integración
-
Memoria core en el contexto
-
Varía
-
Precios de nube
-
No especificado
-
Sin presupuesto de tokens
-
Respaldado por LLM (varía)
-
Varía
-
22K+ tokens con 240 obs
-
-
-
Visor en tiempo real
-
Sí (puerto 3113)
-
Dashboard en la nube
-
Dashboard en la nube
-
UI web
-
Dashboard en la nube
-
UI web del hub
-
No
-
No
-
No
-
No
-
-
-
Self-hosted
-
Sí (por defecto)
-
Opcional
-
Opcional
-
Sí
-
No (solo nube)
-
Sí (Docker)
-
Sí
-
Sí (Oracle DB)
-
Sí
-
Sí
-
-
-
-Nota de benchmark: solo el R@5 de agentmemory es un resultado medido por nosotros (LongMemEval-S, reproducible desde benchmark/COMPARISON.md). Las cifras de mem0 y Letta son sus números publicados de LoCoMo (un dataset distinto); las cifras de MemPalace, supermemory, TencentDB (PersonaMem) y oracleagentmemory son afirmaciones autoinformadas por el proveedor que no hemos reproducido de forma independiente (la ejecución de oracleagentmemory usó GPT-5.5 contra una Oracle AI Database). Se muestran una junto a otra solo a modo orientativo, no como un cara a cara sobre datos idénticos. Los conteos de estrellas son aproximados y varían con el tiempo.
-
-**Nuevos entrantes** que vale la pena conocer, comparados en profundidad en [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Sistema | ⭐ | Enfoque |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Grafo de conocimiento temporal; los resultados publicados más fuertes en consultas temporales (LongMemEval 63.8%), pero el grafo se construye de forma asíncrona, así que los hechos recientes pueden tardar en aparecer |
-| Cognee | 30K | Ingesta de documento a grafo de conocimiento, solo Python, construido para extracción estructurada de entidades en lugar de captura de sesiones |
+Comparación actualizada de productos y opciones de despliegue, con fuentes:
-AgentMemory combina captura desde agentes de programación, historial de proyecto consultable y un visor local. Puedes usarlo sin clave desde el primer inicio.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -779,7 +571,7 @@ Guía completa: [`integrations/hermes/`](../integrations/hermes/)
Arranca el servidor de memoria: `npx -y @agentmemory/agentmemory@latest`
-#### Skills nativas vía `npx skills add` (50+ agentes)
+#### Skills nativas vía `npx skills add`
agentmemory incluye 17 skills en el formato `/SKILL.md` al estilo de Claude Code: 9 skills de acción invocables (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) y 8 skills de referencia que el agente carga bajo demanda (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Las skills de referencia llevan tablas de datos generadas desde el código fuente, así que nunca se desincronizan. La CLI [`skills`](https://npmjs.com/package/skills) de vercel-labs las auto-instala en el directorio de skills nativo del agente que la invoca en 50+ agentes (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf y más):
@@ -836,7 +628,7 @@ La entrada de agentmemory es el **mismo bloque de servidor MCP** en todos los ho
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` instala el MCP y los hooks de captura en el directorio de personalización compartido. Consulta [configuración y límites de Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` usa la misma configuración de MCP y hooks que las versiones actuales del IDE. Las instalaciones existentes deben actualizar con `--force`; consulta las [notas de actualización](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` escribe la configuración a nivel de usuario. Los overrides de workspace van en `.kiro/settings/mcp.json` junto a tu código. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` escribe el bloque `mcpServers` estándar. Warp también auto-descubre skills desde `.claude/skills/`; una vez instalado el plugin de Claude Code, las 8 skills de agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) aparecen de forma nativa en la paleta de slash commands de Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` escribe el bloque `mcpServers` estándar. Warp también auto-descubre skills desde `.claude/skills/`; una vez instalado el plugin de Claude Code, las 9 skills de agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) aparecen de forma nativa en la paleta de slash commands de Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` escribe el bloque `mcpServers` estándar. Usuarios de la extensión de VS Code: pega el mismo bloque vía Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (preferido) o `config.json` (legacy) | `agentmemory connect continue` crea `config.yaml` si falta o modifica el `config.json` existente. Si ya existe YAML, muestra el bloque que debes añadir a `mcpServers:` para conservar comentarios y anclas. Continue usa un array, no un objeto. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` escribe bajo `context_servers` (la clave de Zed, NO `mcpServers`). Los servidores MCP remotos se pueden conectar en su lugar vía `{"url": "..."}`. |
@@ -927,7 +719,6 @@ Todo agente de codificación con IA viene con memoria integrada: Claude Code tie
|---|---|---|
| Mantenimiento | Instrucciones mantenidas manualmente | Historial de sesiones y recuerdos indexados |
| Búsqueda | Carga todo en contexto | BM25 + vector + graph (solo top-K) |
-| Coste en tokens | 22K+ con 240 observaciones | ~1,900 tokens (92% menos) |
| Cross-agent | Ficheros por agente | MCP + REST (cualquier agente) |
| Coordinación | Ninguna | Leases, signals, actions, routines |
| Observabilidad | Lectura manual de ficheros | Visor en tiempo real en :3113 |
@@ -940,7 +731,7 @@ Todo agente de codificación con IA viene con memoria integrada: Claude Code tie
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1025,7 +816,7 @@ Fusionado con Reciprocal Rank Fusion (RRF, k=60) y diversificado por sesión (m
Cuando el índice vectorial está poblado, `mem::search` (detrás de `memory_recall`) usa el ranker híbrido BM25 + vector. Sin embeddings usa BM25. `smart-search` además puede fusionar coincidencias estructurales del grafo cuando existen datos de grafo, incluso en modo sin claves. El recall de lecciones corre sobre un índice BM25 in-memory dedicado en lugar de escanear todo el corpus en cada consulta. Las versiones de memoria reemplazadas quedan excluidas de todas las rutas de recall; la cadena de versiones conserva su historial.
-Los vectores sobreviven a un crash o un force-kill. El índice vectorial se guarda en buckets como máximo cada `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minutos). Cada vector añadido o eliminado entretanto también se escribe de inmediato en un pequeño log pendiente en el state store, y el siguiente arranque lo reproduce sin llamar al proveedor de embeddings. Cada guardado exitoso vacía el log. Los documentos que siguen sin vector tras la reproducción se re-embeben en segundo plano en lotes de `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) hasta que no quede ninguno, y un backfill que se detiene continúa en el siguiente arranque. `/agentmemory/status` y el visor muestran el tamaño del log pendiente y el estado del backfill. Las instalaciones sin claves no escriben nada.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokeniza griego, cirílico, hebreo, árabe y latín con tildes de serie. Para memorias en chino / japonés / coreano, instala los segmentadores opcionales (`npm install @node-rs/jieba tiny-segmenter`) para partir los runs CJK en tokens a nivel de palabra; sin ellos, agentmemory hace soft-fallback a tokenización por run completo y muestra una pista única en stderr.
@@ -1041,9 +832,9 @@ La instalación npm normal incluye el runtime opcional `@huggingface/transformer
| Proveedor | Modelo | Coste | Notas |
|---|---|---|---|
-| **Local (opt-in recomendado)** | `all-MiniLM-L6-v2` | Gratis | En local tras la primera descarga del modelo, +8pp de recall sobre BM25-only |
+| **Local (opt-in recomendado)** | `all-MiniLM-L6-v2` | Gratis | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Free tier | 100+ idiomas, 768/1536/3072 dims (MRL), entrada de 2048 tokens. Sustituye a `text-embedding-004` ([deprecado, cierre el 14 ene 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Máxima calidad |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | De pago | Optimizado para código |
| Cohere | `embed-english-v3.0` | Trial gratis | Uso general |
| OpenRouter | Cualquier modelo | Varía | Proxy multi-modelo |
diff --git a/READMEs/README.fi-FI.md b/READMEs/README.fi-FI.md
index 02dab56fdb..24c7d17060 100644
--- a/READMEs/README.fi-FI.md
+++ b/READMEs/README.fi-FI.md
@@ -66,8 +66,6 @@
-
-Benchmark-huomautus: vain agentmemoryn R@5 on omaa mitattua tulostamme (LongMemEval-S, toistettavissa tiedostosta benchmark/COMPARISON.md). mem0- ja Letta-luvut ovat niiden julkaisemia LoCoMo-lukuja (eri aineisto); MemPalace-, supermemory-, TencentDB (PersonaMem) - ja oracleagentmemory-luvut ovat valmistajien itse ilmoittamia väitteitä, joita emme ole itsenäisesti toistaneet (oracleagentmemoryn ajo käytti GPT-5.5:tä Oracle AI Databasea vastaan). Esitetty rinnakkain vain suuntaa antavana, ei pää-pää-vertailuna samalla datalla. Tähtimäärät ovat likimääräisiä ja muuttuvat ajan myötä.
-
-**Uudemmat tulokkaat**, hyvä tietää, syvemmin vertailtuna kohdassa [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Järjestelmä | ⭐ | Näkökulma |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporaalinen tietograafi; vahvimmat julkaistut temporaalikyselytulokset (LongMemEval 63.8%), mutta graafi rakentuu asynkronisesti, jolloin tuoreet faktat voivat viivästyä |
-| Cognee | 30K | Dokumentista tietograafiksi -tuonti, vain Python, rakennettu rakenteelliseen entiteettipoimintaan istuntotallennuksen sijaan |
+Päivitetty tuotteiden ja käyttöönottovaihtoehtojen vertailu lähteineen:
-AgentMemory yhdistää istuntojen tallennuksen, haettavan projektihistorian ja paikallisen viewerin; avaimeton käyttö toimii ensimmäisestä käynnistyksestä.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -786,7 +578,7 @@ Täysi opas: [`integrations/hermes/`](../integrations/hermes/)
Käynnistä muistipalvelin: `npx -y @agentmemory/agentmemory@latest`
-#### Natiivit taidot komennolla `npx skills add` (50+ agenttia)
+#### Natiivit taidot komennolla `npx skills add`
agentmemory toimittaa 17 taitoa Claude-Code-tyylisessä `/SKILL.md`-formaatissa: 9 kutsuttavaa toimintotaitoa (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) ja 8 viitetaitoa, jotka agentti lataa tarpeen mukaan (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Viitetaidot kuljettavat lähdekoodista generoituja datataulukoita, niin ne eivät koskaan ajaudu pois synkasta. vercel-labsin [`skills`](https://npmjs.com/package/skills)-CLI asentaa ne automaattisesti kutsuvan agentin natiiviin taitohakemistoon yli 50 agentin yli (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf ja monet muut):
@@ -844,7 +636,7 @@ agentmemory-merkintä on **sama MCP-palvelinlohko** jokaisessa isännässä, jok
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` asentaa MCP:n ja tallennushookit jaettuun mukautushakemistoon. Katso [Antigravityn asennus ja rajoitukset](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` käyttää samaa MCP- ja hook-määritystä kuin nykyiset IDE-versiot. Olemassa olevat asennukset tulisi päivittää `--force`-valitsimella; katso [päivitysohjeet](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` kirjoittaa käyttäjätason asetukset. Työtilakohtaiset ylikirjoitukset menevät tiedostoon `.kiro/settings/mcp.json` koodisi vierelle. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` kirjoittaa standardin `mcpServers`-lohkon. Warp myös tunnistaa automaattisesti taidot kansiosta `.claude/skills/`; kun Claude Code -plugin on asennettu, 8 agentmemory-taitoa (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) näkyvät natiivisti Warpin pikakomentopaletissa. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` kirjoittaa standardin `mcpServers`-lohkon. Warp myös tunnistaa automaattisesti taidot kansiosta `.claude/skills/`; kun Claude Code -plugin on asennettu, 9 agentmemory-taitoa (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) näkyvät natiivisti Warpin pikakomentopaletissa. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` kirjoittaa standardin `mcpServers`-lohkon. VS Code -laajennuksen käyttäjät: liitä sama lohko Cline Settings → MCP Servers → Edit JSON -kautta. |
| **Continue.dev** | `~/.continue/config.yaml` (suositeltu) tai `config.json` (vanha) | `agentmemory connect continue` luo `config.yaml`-tiedoston, jos asetuksia ei ole, tai muokkaa olemassa olevaa `config.json`-tiedostoa. Olemassa olevan YAML:n tapauksessa se tulostaa lisättävän `mcpServers:`-lohkon säilyttääkseen kommentit ja ankkurit. Continue käyttää taulukkoa, ei objektia. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` kirjoittaa avaimen `context_servers` alle (Zedin avain, EI `mcpServers`). Etä-MCP-palvelimet voidaan kytkeä myös muodolla `{"url": "..."}`. |
@@ -927,7 +719,6 @@ Jokainen tekoälypohjainen koodausagentti toimitetaan sisäänrakennetulla muist
|---|---|---|
| Ylläpito | Käsin ylläpidetyt ohjeet | Indeksoitu istuntohistoria ja muistot |
| Haku | Lataa kaiken kontekstiin | BM25 + vektori + graafi (vain top-K) |
-| Tokenkustannus | 22K+ 240 havainnolla | ~1 900 tokenia (92 % vähemmän) |
| Agenttien välillä | Agenttikohtaiset tiedostot | MCP + REST (mikä tahansa agentti) |
| Koordinointi | Ei mitään | Leaset, signaalit, toiminnot, rutiinit |
| Havainnoitavuus | Lue tiedostoja käsin | Reaaliaikainen katseluohjelma portissa :3113 |
@@ -940,7 +731,7 @@ Jokainen tekoälypohjainen koodausagentti toimitetaan sisäänrakennetulla muist
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1025,7 +816,7 @@ Yhdistetty Reciprocal Rank Fusionilla (RRF, k=60) ja istuntokohtaisesti monipuol
Kun vektori-indeksi on täytetty, `mem::search` (`memory_recall`-työkalun takana) käyttää hybridiä BM25 + vektori -luokittelijaa. Ilman upotuksia se käyttää BM25:tä. `smart-search` voi lisäksi yhdistää rakenteellisia graafiosumia, kun graafidataa on olemassa, myös avaimettomassa tilassa. Opetusten palautus käyttää erillistä muistinvaraista BM25-indeksiä sen sijaan, että se skannaisi koko korpuksen jokaisella kyselyllä. Korvatut muistiversiot on suljettu pois jokaisesta palautusreitistä; versioketju säilyttää niiden historian.
-Vektorit selviävät kaatumisesta tai pakkosammutuksesta. Vektori-indeksi tallennetaan erissä enintään `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS`-ajan välein (10 minuuttia). Joka vektori, joka lisätään tai poistetaan tällä välillä, kirjoitetaan myös heti pieneen odottavaan lokiin tilavarastossa, ja seuraava käynnistys toistaa sen kutsumatta upotustarjoajaa. Joka onnistunut tallennus tyhjentää lokin. Dokumentit, joilla ei vielä ole vektoria toiston jälkeen, upotetaan uudelleen taustalla erissä kooltaan `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500), kunnes yhtäkään ei jää jäljelle, ja pysäytetty täydennys jatkuu seuraavassa käynnistyksessä. `/agentmemory/status` ja katseluohjelma näyttävät odottavan lokin koon ja täydennyksen tilan. Avaimettomat asennukset eivät kirjoita mitään.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokenisoi kreikan, kyrilliset kirjaimet, hebrean, arabian ja aksentoidun latinan oletuksena. Kiinan-, japanin- ja koreankielisiä muistoja varten asenna valinnaiset segmentoijat (`npm install @node-rs/jieba tiny-segmenter`) jakaaksesi CJK-jaksot sananlaajuisiksi tokeneiksi; ilman niitä agentmemory lankeaa pehmeästi koko jakson tokenisointiin ja tulostaa kertaluonteisen vihjeen stderriin.
@@ -1041,9 +832,9 @@ Normaali npm-asennus sisältää valinnaisen `@huggingface/transformers`-ajonaik
| Tarjoaja | Malli | Kustannus | Huomautuksia |
|---|---|---|---|
-| **Paikallinen (suositeltu opt-in)** | `all-MiniLM-L6-v2` | Ilmainen | Laitteella ensimmäisen mallilatauksen jälkeen, +8pp palautustarkkuutta vain-BM25:een verrattuna |
+| **Paikallinen (suositeltu opt-in)** | `all-MiniLM-L6-v2` | Ilmainen | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Ilmainen taso | 100+ kieltä, 768/1536/3072-ulotteinen (MRL), 2048-tokenin syöte. Korvaa `text-embedding-004`:n ([poistettu käytöstä, sammutetaan 14.1.2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | 0,02 $/1M | Korkein laatu |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Maksullinen | Optimoitu koodille |
| Cohere | `embed-english-v3.0` | Ilmainen kokeilu | Yleiskäyttöinen |
| OpenRouter | Mikä tahansa malli | Vaihtelee | Multi-model-proxy |
diff --git a/READMEs/README.fr-FR.md b/READMEs/README.fr-FR.md
index c0ef2f9df1..e7172efac7 100644
--- a/READMEs/README.fr-FR.md
+++ b/READMEs/README.fr-FR.md
@@ -66,8 +66,6 @@
-
-
@@ -331,223 +329,17 @@ Notes de la dernière version : [CHANGELOG.md](../CHANGELOG.md).
-
-
-
-
-### Précision de récupération
-
-**coding-agent-life-v1** (corpus interne, reproductible en sandbox)
-
-| Adaptateur | P@5 | R@5 | Taux de hit top-5 | Latence p50 |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| Référence grep | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-100% de taux de hit top-5 au **plafond mathématique du P@5** pour ce corpus (0.240, voir le scorecard). L'hybride récupère chaque session gold ; grep manque 1 des 2 gold sur la requête temporelle multi-session. Le gain porte sur **recall + temporel**, pas sur la précision agrégée. Ce benchmark est petit et pauvre en gold ; le LongMemEval-S plus grand ci-dessous différencie mieux. Ventilation complète par type + note de correction : [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 questions)
+Résultats actuels de recherche, définitions des métriques et étapes de reproduction :
-| Système | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| Repli BM25 seul | 86.2% | 94.6% | 71.5% |
-
-
-
-> Modèle d'embedding : `all-MiniLM-L6-v2` (local, gratuit, sans clé d'API). Rapports complets : [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Comparaison avec les concurrents : [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) couvrant agentmemory vs mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Reproduire localement :** [`eval/README.md`](../eval/README.md), un harnais à adaptateurs pluggables pour LongMemEval `_s` (public, 500 questions) + `coding-agent-life-v1` (corpus interne de 15 sessions). Les adaptateurs grep / vectoriel / agentmemory sont notés côte à côte, sortie NDJSON, scorecards publiés dans [`docs/benchmarks/`](../docs/benchmarks/).
-
-**À associer à [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) et [Graphify](https://github.com/safishamsi/graphify).** Indexation de graphe de code, pipelines de build multi-agents et graphes de connaissances plus larges sur docs / PDF / images / vidéos. agentmemory mémorise le travail ; ces trois projets éclairent le reste de la couche de contexte. Recettes + tableau de routage des questions : [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Intégré (CLAUDE.md)
-
-
-
Type
-
Moteur de mémoire + serveur MCP
-
API de couche mémoire
-
Runtime d'agent complet
-
IA personnelle
-
API mémoire + app
-
Hub de mémoire d'équipe (proxy LLM)
-
Mémoire vectorielle (OSS)
-
Moteur de mémoire (Oracle DB)
-
Système de mémoire
-
Fichier statique
-
-
-
R@5 de récupération
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Auto-déclaré
-
PersonaMem 76% (auto-déclaré)
-
~96.6% (auto-déclaré)
-
94.4% (auto-déclaré)
-
N/A
-
N/A (grep)
-
-
-
Capture automatique
-
12 hooks (zéro effort manuel)
-
Appels add() manuels
-
L'agent s'auto-édite
-
Manuelle
-
Extraction côté API
-
Interception par proxy (bascule de base-URL)
-
Manuelle
-
Extraction API
-
Manuelle
-
Édition manuelle
-
-
-
Recherche
-
BM25 + Vectoriel + Graphe (fusion RRF)
-
Vectoriel + Graphe
-
Vectoriel (archival)
-
Sémantique
-
Vectoriel + RAG
-
4 types d'assets (Chat / Skill / Wiki / CodeGraph)
-
Vectoriel uniquement
-
Vectoriel + sémantique
-
Pondérée par décroissance
-
Charge tout en contexte
-
-
-
Multi-agents
-
MCP + REST + leases + signaux
-
API (sans coordination)
-
Uniquement dans le runtime Letta
-
Non
-
Non
-
Rôles d'équipe + assets partagés
-
Non
-
Scopé seulement
-
Partagé multi-agents
-
Fichiers par agent
-
-
-
Verrouillage framework
-
Aucun (tout client MCP)
-
Aucun
-
Élevé (Letta obligatoire)
-
Autonome
-
Aucun
-
Le proxy s'interpose devant chaque appel de modèle
-
Aucun
-
Oracle Database
-
Aucun
-
Format par agent
-
-
-
Dépendances externes
-
Aucun service de base de données séparé requis
-
Qdrant / pgvector
-
Postgres + base vectorielle
-
Multiples
-
Cloud managé
-
Stack Docker (Core + Hub + Proxy)
-
Store vectoriel
-
Oracle AI Database
-
Aucune
-
Aucune
-
-
-
Cycle de vie mémoire
-
Consolidation à 4 niveaux + décroissance + oubli automatique
-
Extraction passive
-
Géré par l'agent
-
Manuel
-
Oubli automatique
-
Revue manuelle ; auto-routage en cours
-
Aucun
-
Non précisé
-
Décroissance + consolidation
-
Nettoyage manuel
-
-
-
Efficacité des tokens
-
~1 900 tokens/session (10 $/an)
-
Varie selon l'intégration
-
Mémoire centrale dans le contexte
-
Varie
-
Tarification cloud
-
Non précisé
-
Pas de budget de tokens
-
Adossé à un LLM (variable)
-
Varie
-
22K+ tokens à 240 observations
-
-
-
Visualiseur temps réel
-
Oui (port 3113)
-
Dashboard cloud
-
Dashboard cloud
-
UI web
-
Dashboard cloud
-
UI web du Hub
-
Non
-
Non
-
Non
-
Non
-
-
-
Auto-hébergé
-
Oui (par défaut)
-
Optionnel
-
Optionnel
-
Oui
-
Non (cloud uniquement)
-
Oui (Docker)
-
Oui
-
Oui (Oracle DB)
-
Oui
-
Oui
-
-
-
-Note sur le benchmark : seul le R@5 d'agentmemory est notre propre résultat mesuré (LongMemEval-S, reproductible depuis benchmark/COMPARISON.md). Les chiffres de mem0 et Letta sont leurs résultats LoCoMo publiés (un jeu de données différent) ; les chiffres de MemPalace, supermemory, TencentDB (PersonaMem) et oracleagentmemory sont des annonces auto-déclarées par ces éditeurs, que nous n'avons pas reproduites indépendamment (le run d'oracleagentmemory utilisait GPT-5.5 contre une Oracle AI Database). Présentés côte à côte à titre indicatif seulement, pas comme un comparatif direct sur des données identiques. Les comptages d'étoiles sont approximatifs et évoluent avec le temps.
-
-**Nouveaux arrivants** à connaître, comparés en détail dans [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) :
-
-| Système | ⭐ | Angle |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Graphe de connaissances temporel ; meilleurs résultats publiés sur les requêtes temporelles (LongMemEval 63.8%), mais le graphe se construit de façon asynchrone donc les faits récents peuvent accuser un retard |
-| Cognee | 30K | Ingestion document-vers-graphe-de-connaissances, Python uniquement, conçu pour l'extraction d'entités structurées plutôt que pour la capture de sessions |
+Comparaison actualisée des produits et des options de déploiement, avec sources :
-AgentMemory associe capture par les agents de programmation, historique de projet interrogeable et visualiseur local. Le fonctionnement sans clé est disponible dès le premier lancement.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -779,7 +571,7 @@ Guide complet : [`integrations/hermes/`](../integrations/hermes/)
Démarrez le serveur de mémoire : `npx -y @agentmemory/agentmemory@latest`
-#### Skills natifs via `npx skills add` (50+ agents)
+#### Skills natifs via `npx skills add`
agentmemory livre 17 skills au format `/SKILL.md` façon Claude Code : 9 skills d'action invocables (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) et 8 skills de référence que l'agent charge à la demande (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Les skills de référence embarquent des tableaux de données générés depuis les sources, donc ils ne dérivent jamais. La CLI [`skills`](https://npmjs.com/package/skills) de vercel-labs les installe automatiquement dans le répertoire de skills natif de l'agent appelant sur 50+ agents (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, et plus) :
@@ -836,7 +628,7 @@ L'entrée agentmemory est le **même bloc de serveur MCP** sur tous les hôtes q
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` installe MCP et les hooks de capture dans le répertoire de personnalisation partagé. Voir [Configuration et limites d'Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` utilise la même configuration MCP et hooks que les versions actuelles de l'IDE. Les installations existantes doivent se rafraîchir avec `--force` ; voir les [notes de mise à jour](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` écrit la config au niveau utilisateur. Les surcharges au niveau du workspace vont dans `.kiro/settings/mcp.json` à côté de votre code. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` écrit le bloc `mcpServers` standard. Warp découvre aussi automatiquement les skills depuis `.claude/skills/` ; une fois le plugin Claude Code installé, les 8 skills agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) apparaissent nativement dans la palette de commandes slash de Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` écrit le bloc `mcpServers` standard. Warp découvre aussi automatiquement les skills depuis `.claude/skills/` ; une fois le plugin Claude Code installé, les 9 skills agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) apparaissent nativement dans la palette de commandes slash de Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` écrit le bloc `mcpServers` standard. Utilisateurs de l'extension VS Code : collez le même bloc via Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (préféré) ou `config.json` (legacy) | `agentmemory connect continue` crée une `config.yaml` absente ou modifie la `config.json` existante. Si le YAML existe, l'adaptateur affiche le bloc à ajouter sous `mcpServers:` pour préserver commentaires et ancres. Continue attend un tableau, pas un objet. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` écrit sous `context_servers` (la clé de Zed, PAS `mcpServers`). Les serveurs MCP distants peuvent être câblés via `{"url": "..."}` à la place. |
@@ -934,7 +726,6 @@ Chaque agent de codage IA est livré avec une mémoire intégrée : Claude Code
|---|---|---|
| Entretien | Instructions maintenues manuellement | Historique des sessions et souvenirs indexés |
| Recherche | Charge tout en contexte | BM25 + vectoriel + graphe (top-K seulement) |
-| Coût en tokens | 22K+ à 240 observations | ~1 900 tokens (92% de moins) |
| Entre agents | Fichiers par agent | MCP + REST (tout agent) |
| Coordination | Aucune | Leases, signaux, actions, routines |
| Observabilité | Lire les fichiers manuellement | Visualiseur temps réel sur :3113 |
@@ -947,7 +738,7 @@ Chaque agent de codage IA est livré avec une mémoire intégrée : Claude Code
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1032,7 +823,7 @@ Fusionné avec la Reciprocal Rank Fusion (RRF, k=60) et diversifié par session
Quand un index vectoriel est peuplé, `mem::search` (derrière `memory_recall`) utilise le ranker hybride BM25 + vectoriel. Sans embeddings, il utilise BM25. `smart-search` peut en plus fusionner des correspondances de graphe structurel quand des données de graphe existent, y compris en mode sans clé. Le recall de lessons tourne sur un index BM25 en mémoire dédié plutôt que de scanner tout le corpus à chaque requête. Les versions de mémoire supersédées sont exclues de tout chemin de recall ; la chaîne de versions conserve leur historique.
-Les vecteurs survivent à un crash ou à un force-kill. L'index vectoriel est sauvegardé par lots au plus toutes les `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minutes). Chaque vecteur ajouté ou retiré entre-temps est aussi écrit immédiatement dans un petit journal en attente dans le state store, et le prochain démarrage le rejoue sans appeler le fournisseur d'embedding. Chaque sauvegarde réussie vide le journal. Les documents qui n'ont toujours pas de vecteur après le replay sont réembarqués en arrière-plan par lots de `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) jusqu'à épuisement, et un backfill arrêté reprend au démarrage suivant. `/agentmemory/status` et le visualiseur affichent la taille du journal en attente et l'état du backfill. Les installations sans clé n'écrivent rien.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokenise le grec, le cyrillique, l'hébreu, l'arabe et le latin accentué nativement. Pour les mémoires en chinois / japonais / coréen, installez les segmenteurs optionnels (`npm install @node-rs/jieba tiny-segmenter`) pour découper les suites CJK en tokens au niveau mot ; sans eux, agentmemory retombe en douceur sur une tokenisation en bloc complet et affiche un indice ponctuel sur stderr.
@@ -1048,9 +839,9 @@ L'installation npm normale inclut le runtime optionnel `@huggingface/transformer
| Fournisseur | Modèle | Coût | Notes |
|---|---|---|---|
-| **Local (opt-in recommandé)** | `all-MiniLM-L6-v2` | Gratuit | Sur l'appareil après le premier téléchargement de modèle, +8pp de recall par rapport à BM25 seul |
+| **Local (opt-in recommandé)** | `all-MiniLM-L6-v2` | Gratuit | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Palier gratuit | 100+ langues, 768/1536/3072 dims (MRL), entrée de 2048 tokens. Remplace `text-embedding-004` ([obsolète, arrêt le 14 janvier 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | 0.02 $/1M | Meilleure qualité |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Payant | Optimisé pour le code |
| Cohere | `embed-english-v3.0` | Essai gratuit | Usage général |
| OpenRouter | N'importe quel modèle | Variable | Proxy multi-modèle |
diff --git a/READMEs/README.he-IL.md b/READMEs/README.he-IL.md
index 4eb23cb2d5..e18522b845 100644
--- a/READMEs/README.he-IL.md
+++ b/READMEs/README.he-IL.md
@@ -68,8 +68,6 @@
-
-### דיוק שליפה
-
-**coding-agent-life-v1** (קורפוס פנימי, שניתן לשכפול ב-sandbox)
-
-| מתאם | P@5 | R@5 | שיעור פגיעה ב-Top-5 | latency p50 |
-|---|---|---|---|---|
-| **agentmemory היברידי** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep (קו בסיס) | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-שיעור פגיעה של 100% ב-top-5 בתקרה המתמטית של **P@5** עבור הקורפוס הזה (0.240, ראו scorecard). המצב ההיברידי שולף כל session זהב (gold); ל-grep חסר 1 מתוך 2 פריטי gold בשאילתה הטמפורלית המשתרעת על כמה sessions. השיפור הוא ב-**recall + טמפורליות**, לא בדיוק מצרפי. ה-benchmark הזה קטן ומעוט gold; LongMemEval-S הגדול יותר שלהלן מבדיל טוב יותר. פירוט מלא לפי טיפוס + הערת תיקון: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 שאלות)
+תוצאות האחזור העדכניות, הגדרות המדדים ושלבי השחזור:
-| מערכת | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| חלופת BM25 בלבד | 86.2% | 94.6% | 71.5% |
-
-
-
-
-### חיסכון בטוקנים
-
-| גישה | טוקנים/שנה | עלות/שנה |
-|---|---|---|
-| הדבקת הקשר מלא | 19.5M+ | בלתי אפשרי (חורג מהחלון) |
-| סיכום על ידי LLM | ~650K | ~$500 |
-| **agentmemory** | **~170K** | **~$10** |
-| agentmemory + embeddings מקומיים | ~170K | **$0** |
-
-
-
-
-
-> מודל embedding: `all-MiniLM-L6-v2` (מקומי, חינמי, בלי מפתח API). דוחות מלאים: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). השוואת מתחרים: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) המכסה את agentmemory לעומת mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**שכפול מקומי:** [`eval/README.md`](../eval/README.md), harness עם adapters נתיקים עבור LongMemEval `_s` (500 שאלות ציבוריות) + `coding-agent-life-v1` (קורפוס פנימי של 15 sessions). מתאמי Grep / וקטור / agentmemory מקבלים ניקוד אחד לצד השני, פלט NDJSON, וה-scorecards המתפרסמים נוחתים ב-[`docs/benchmarks/`](../docs/benchmarks/).
-
-**משתלב עם [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything), ו-[Graphify](https://github.com/safishamsi/graphify).** אינדוקס code-graph, pipelines לבנייה מרובת-סוכנים, וגרפי ידע רחבים יותר על פני docs / PDF / תמונות / וידאו. agentmemory זוכר את העבודה; שלושת הפרויקטים האלה מדליקים את שאר שכבת ההקשר. מתכונים + טבלת ניתוב שאלות: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Built-in (CLAUDE.md)
-
-
-
סוג
-
מנוע זיכרון + שרת MCP
-
API של שכבת זיכרון
-
runtime מלא לסוכן
-
AI אישי
-
API זיכרון + אפליקציה
-
hub זיכרון לצוות (proxy ל-LLM)
-
זיכרון וקטורי (קוד פתוח)
-
מנוע זיכרון (Oracle DB)
-
מערכת זיכרון
-
קובץ סטטי
-
-
-
שליפה R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
דיווח עצמי
-
PersonaMem 76% (דיווח עצמי)
-
~96.6% (דיווח עצמי)
-
94.4% (דיווח עצמי)
-
N/A
-
N/A (grep)
-
-
-
לכידה אוטומטית
-
12 hooks (ללא מאמץ ידני)
-
קריאות add() ידניות
-
עריכות עצמיות של הסוכן
-
ידני
-
חילוץ בצד ה-API
-
יירוט proxy (החלפת base-URL)
-
ידני
-
חילוץ API
-
ידני
-
עריכה ידנית
-
-
-
חיפוש
-
BM25 + וקטור + גרף (fusion מסוג RRF)
-
וקטור + גרף
-
וקטור (ארכיוני)
-
סמנטי
-
וקטור + RAG
-
4 סוגי נכסים (Chat / Skill / Wiki / CodeGraph)
-
וקטור בלבד
-
וקטור + סמנטי
-
משוקלל decay
-
טוען הכול להקשר
-
-
-
ריבוי-סוכנים
-
MCP + REST + leases + signals
-
API (בלי תיאום)
-
רק בתוך runtime של Letta
-
לא
-
לא
-
תפקידי צוות + נכסים משותפים
-
לא
-
מוגבל לתחום בלבד
-
משותף בין סוכנים מרובים
-
קבצים לפי סוכן
-
-
-
נעילה ל-framework
-
אין (כל לקוח MCP)
-
אין
-
גבוהה (חייבים להשתמש ב-Letta)
-
עצמאי (standalone)
-
אין
-
Proxy חוצץ בפני כל קריאת מודל
-
אין
-
Oracle Database
-
אין
-
פורמט לפי סוכן
-
-
-
תלויות חיצוניות
-
אין צורך בשירות מסד נתונים נפרד
-
Qdrant / pgvector
-
Postgres + מסד נתונים וקטורי
-
מרובות
-
cloud מנוהל
-
ערימת Docker (Core + Hub + Proxy)
-
מאגר וקטורי
-
Oracle AI Database
-
אין
-
אין
-
-
-
מחזור חיים של הזיכרון
-
קונסולידציה ב-4 שכבות + decay + שכחה אוטומטית
-
חילוץ פסיבי
-
מנוהל על ידי הסוכן
-
ידני
-
שכחה אוטומטית
-
בדיקה ידנית; ניתוב אוטומטי בפיתוח
-
אין
-
לא מצוין
-
decay + קונסולידציה
-
גיזום ידני
-
-
-
יעילות טוקנים
-
~1,900 טוקנים/session ($10/שנה)
-
משתנה לפי אינטגרציה
-
זיכרון core בתוך ההקשר
-
משתנה
-
תמחור cloud
-
לא מצוין
-
אין תקציב טוקנים
-
מגובה LLM (משתנה)
-
משתנה
-
22K+ טוקנים ב-240 תצפיות
-
-
-
מציג בזמן אמת
-
כן (פורט 3113)
-
dashboard ב-cloud
-
dashboard ב-cloud
-
UI מבוסס web
-
dashboard ב-cloud
-
UI web של ה-hub
-
לא
-
לא
-
לא
-
לא
-
-
-
אירוח עצמי
-
כן (ברירת מחדל)
-
אופציונלי
-
אופציונלי
-
כן
-
לא (cloud בלבד)
-
כן (Docker)
-
כן
-
כן (Oracle DB)
-
כן
-
כן
-
-
-
-הערת benchmark: רק ה-R@5 של agentmemory הוא תוצאת מדידה שלנו (LongMemEval-S, ניתנת לשכפול מ-benchmark/COMPARISON.md). המספרים של mem0 ו-Letta הם המספרים המתפרסמים שלהם מ-LoCoMo (dataset שונה); המספרים של MemPalace, supermemory, TencentDB (PersonaMem), ו-oracleagentmemory הם טענות שדווחו על ידי היצרן ולא שכפלנו אותן באופן עצמאי (ההרצה של oracleagentmemory השתמשה ב-GPT-5.5 מול Oracle AI Database). מוצגים זה לצד זה להערכה כללית בלבד, לא כהשוואת ראש-בראש על אותם נתונים. ספירות הכוכבים הן משוערות ומשתנות עם הזמן.
-
-**שחקנים חדשים יותר** שכדאי להכיר, מושווים בהרחבה ב-[`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| מערכת | ⭐ | זווית |
-|--------|---|-------|
-| Zep / Graphiti | 30K | גרף ידע טמפורלי; תוצאות השאילתה הטמפורלית הפרסומיות החזקות ביותר (LongMemEval 63.8%), אבל הגרף נבנה באופן אסינכרוני כך שעובדות חדשות יכולות להתעכב |
-| Cognee | 30K | בליעת מסמכים לגרף ידע (document-to-knowledge-graph), Python בלבד, נבנה לחילוץ ישויות מבני ולא ללכידת sessions |
+השוואה מעודכנת של מוצרים ואפשרויות פריסה, עם מקורות:
-AgentMemory משלב תיעוד סשנים, היסטוריית פרויקט שניתן לחפש בה וממשק צפייה מקומי, עם עבודה ללא מפתח כבר מההפעלה הראשונה.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -788,7 +580,7 @@ Verify with `curl http://localhost:3111/agentmemory/health`. Open http://localho
הפעילו את שרת הזיכרון: `npx -y @agentmemory/agentmemory@latest`
-#### Skills מובנים דרך `npx skills add` (50+ סוכנים)
+#### Skills מובנים דרך `npx skills add`
agentmemory שולח 17 skills בפורמט `/SKILL.md` בסטייל Claude Code: 9 skills של פעולה שניתן להפעיל (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) ו-8 skills של עיון שהסוכן טוען לפי הצורך (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). ה-skills של העיון נושאים טבלאות נתונים שנוצרות מהמקור, כך שהן לעולם לא נסחפות (drift). ה-CLI [`skills`](https://npmjs.com/package/skills) של vercel-labs מתקין אותם אוטומטית לתוך תיקיית ה-skills המובנית של הסוכן הקורא על פני 50+ סוכנים (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, ועוד):
@@ -846,7 +638,7 @@ npx skills add rohitg00/agentmemory -y -a '*' # install to every installed age
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` מתקין MCP ו-hooks של לכידה בתיקיית ההתאמה האישית המשותפת. ראו [הגדרה ומגבלות של Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` משתמש באותה תצורת MCP ו-hooks כמו גרסאות ה-IDE הנוכחיות. התקנות קיימות צריכות לרפרש עם `--force`; ראו את [הערות השדרוג](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` כותב את התצורה בשכבת המשתמש. override-ים בשכבת ה-workspace נכנסים ב-`.kiro/settings/mcp.json` לצד הקוד שלכם. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` כותב את בלוק ה-`mcpServers` הסטנדרטי. Warp גם מגלה אוטומטית skills מ-`.claude/skills/`; כשה-plugin של Claude Code מותקן, 8 ה-skills של agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) מופיעים באופן מובנה בפלטת פקודות ה-slash של Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` כותב את בלוק ה-`mcpServers` הסטנדרטי. Warp גם מגלה אוטומטית skills מ-`.claude/skills/`; כשה-plugin של Claude Code מותקן, 9 ה-skills של agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) מופיעים באופן מובנה בפלטת פקודות ה-slash של Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` כותב את בלוק ה-`mcpServers` הסטנדרטי. למשתמשי ה-extension של VS Code: הדביקו את אותו בלוק דרך Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (מועדף) או `config.json` (legacy) | `agentmemory connect continue` יוצר `config.yaml` כשאין הגדרות, או מעדכן `config.json` קיים. עבור YAML קיים הוא מציג בלוק להוספה תחת `mcpServers:`, כדי לשמור הערות ועוגנים. Continue משתמש במערך ולא באובייקט. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` כותב תחת `context_servers` (המפתח של Zed, לא `mcpServers`). שרתי MCP מרוחקים אפשר לחווט במקום זאת דרך `{"url": "..."}`. |
@@ -928,7 +720,6 @@ REST משתמש ב-`3111` וממשק הצפייה ב-`3113`. השאירו אות
|---|---|---|
| תחזוקה | הוראות שנערכות ידנית | היסטוריית סשנים וזיכרונות עם אינדקס |
| חיפוש | טוען הכול להקשר | BM25 + וקטור + גרף (top-K בלבד) |
-| עלות טוקנים | 22K+ ב-240 תצפיות | ~1,900 טוקנים (92% פחות) |
| בין-סוכנים | קבצים לפי סוכן | MCP + REST (כל סוכן) |
| תיאום | אין | leases, signals, actions, routines |
| Observability | קריאת קבצים ידנית | מציג בזמן אמת בפורט :3113 |
@@ -941,7 +732,7 @@ REST משתמש ב-`3111` וממשק הצפייה ב-`3113`. השאירו אות
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1026,7 +817,7 @@ SessionStart hook fires
כשאינדקס וקטורי מאוכלס, `mem::search` (מאחורי `memory_recall`) משתמש ב-ranker ההיברידי של BM25 + וקטור. בלי embeddings הוא משתמש ב-BM25. `smart-search` יכול גם למזג התאמות מבניות מהגרף כשנתוני גרף קיימים, גם במצב keyless. שליפת lessons רצה על אינדקס BM25 ייעודי בזיכרון במקום לסקור את כל הקורפוס בכל שאילתה. גרסאות זיכרון שהוחלפו (superseded) מוחרגות מכל נתיב שליפה; שרשרת הגרסאות שומרת את ההיסטוריה שלהן.
-וקטורים שורדים קריסה או force-kill. אינדקס הווקטורים נשמר ב-buckets לכל היותר כל `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 דקות). כל וקטור שנוסף או הוסר בינתיים גם נכתב מיידית ל-log קטן של pending ב-state store, וההפעלה הבאה משדרת אותו (replay) מחדש בלי לקרוא לספק ה-embedding. כל שמירה מוצלחת מרוקנת את ה-log. מסמכים שעדיין אין להם וקטור אחרי ה-replay מוטבעים (re-embedded) מחדש ברקע ב-batches של `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) עד שלא נשאר אף אחד, ו-backfill שנעצר ממשיך בהפעלה הבאה. `/agentmemory/status` והמציג מציגים את גודל ה-log הממתין ואת מצב ה-backfill. התקנות keyless לא כותבות שום דבר.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 מבצע tokenization ליוונית, קירילית, עברית, ערבית, ולטינית עם סימני הטעמה ישר מהקופסה. לזיכרונות בסינית / יפנית / קוריאנית, התקינו את ה-segmenters האופציונליים (`npm install @node-rs/jieba tiny-segmenter`) כדי לפצל רצפי CJK ל-tokens בגובה-מילה; בלעדיהם, agentmemory נופל בעדינות ל-tokenization של הרצף השלם ומדפיס רמז חד-פעמי ל-stderr.
@@ -1042,9 +833,9 @@ EMBEDDING_PROVIDER=local
| ספק | מודל | עלות | הערות |
|---|---|---|---|
-| **מקומי (opt-in מומלץ)** | `all-MiniLM-L6-v2` | חינמי | במכשיר עצמו אחרי הורדת המודל הראשונה, +8pp ל-recall לעומת BM25 בלבד |
+| **מקומי (opt-in מומלץ)** | `all-MiniLM-L6-v2` | חינמי | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | שכבה חינמית | 100+ שפות, 768/1536/3072 מימדים (MRL), קלט של 2048 טוקנים. מחליף את `text-embedding-004` ([הוצא משימוש, כיבוי ב-14 בינואר 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | האיכות הגבוהה ביותר |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | בתשלום | מותאם לקוד |
| Cohere | `embed-english-v3.0` | ניסיון חינמי | כללי |
| OpenRouter | כל מודל | משתנה | proxy מרובה-מודלים |
diff --git a/READMEs/README.hi-IN.md b/READMEs/README.hi-IN.md
index 5963fa93f8..efee2e6e5c 100644
--- a/READMEs/README.hi-IN.md
+++ b/READMEs/README.hi-IN.md
@@ -66,8 +66,6 @@
-
-> Embedding model: `all-MiniLM-L6-v2` (local, free, कोई API key नहीं)। पूरी रिपोर्ट्स: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md)। प्रतिस्पर्धी तुलना: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md), जो agentmemory बनाम mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo को cover करती है।
-
-**स्थानीय रूप से reproduce करें:** [`eval/README.md`](../eval/README.md), LongMemEval `_s` (public 500-Q) + `coding-agent-life-v1` (in-house 15-session corpus) के लिए एक adapter-pluggable harness। Grep / vector / agentmemory adapters साथ-साथ score होते हैं, NDJSON output, published scorecards [`docs/benchmarks/`](../docs/benchmarks/) में जाते हैं।
-
-**[codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything), और [Graphify](https://github.com/safishamsi/graphify) के साथ जोड़ता है।** Code-graph indexing, multi-agent build pipelines, और docs / PDFs / images / videos में व्यापक knowledge graphs। agentmemory काम याद रखता है; ये तीन projects context layer के बाकी हिस्से को रोशन करते हैं। Recipes + question-routing table: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md)।
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
बिल्ट-इन (CLAUDE.md)
-
-
-
प्रकार
-
Memory engine + MCP सर्वर
-
Memory layer API
-
पूर्ण agent runtime
-
Personal AI
-
Memory API + app
-
Team memory hub (LLM proxy)
-
Vector memory (OSS)
-
Memory engine (Oracle DB)
-
Memory system
-
Static फाइल
-
-
-
Retrieval R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Self-reported
-
PersonaMem 76% (self-reported)
-
~96.6% (self-reported)
-
94.4% (self-reported)
-
N/A
-
N/A (grep)
-
-
-
स्वचालित capture
-
12 hooks (शून्य मैनुअल प्रयास)
-
मैनुअल add() कॉल
-
एजेंट self-edits
-
मैनुअल
-
API-side extraction
-
Proxy interception (base-URL swap)
-
मैनुअल
-
API extraction
-
मैनुअल
-
मैनुअल editing
-
-
-
खोज
-
BM25 + Vector + Graph (RRF fusion)
-
Vector + Graph
-
Vector (archival)
-
Semantic
-
Vector + RAG
-
4 asset types (Chat / Skill / Wiki / CodeGraph)
-
केवल-vector
-
Vector + semantic
-
Decay-weighted
-
सब कुछ context में load करता है
-
-
-
Multi-agent
-
MCP + REST + leases + signals
-
API (कोई coordination नहीं)
-
केवल Letta runtime में
-
नहीं
-
नहीं
-
Team roles + shared assets
-
नहीं
-
केवल scoped
-
Multi-agent shared
-
प्रति-एजेंट फाइलें
-
-
-
Framework lock-in
-
कोई नहीं (कोई भी MCP क्लाइंट)
-
कोई नहीं
-
उच्च (Letta का उपयोग आवश्यक)
-
Standalone
-
कोई नहीं
-
Proxy हर model call के सामने रहता है
-
कोई नहीं
-
Oracle Database
-
कोई नहीं
-
प्रति-एजेंट format
-
-
-
बाहरी निर्भरताएँ
-
अलग डेटाबेस सेवा की जरूरत नहीं
-
Qdrant / pgvector
-
Postgres + vector DB
-
कई
-
Managed cloud
-
Docker stack (Core + Hub + Proxy)
-
Vector store
-
Oracle AI Database
-
कोई नहीं
-
कोई नहीं
-
-
-
Memory lifecycle
-
4-tier consolidation + decay + auto-forget
-
Passive extraction
-
Agent-managed
-
मैनुअल
-
Auto-forget
-
मैनुअल review; auto-routing प्रगति पर
-
कोई नहीं
-
बताया नहीं गया
-
Decay + consolidation
-
मैनुअल pruning
-
-
-
Token दक्षता
-
~1,900 tokens/session ($10/yr)
-
integration पर निर्भर
-
Core memory context में
-
भिन्न
-
Cloud pricing
-
बताया नहीं गया
-
कोई token budget नहीं
-
LLM-backed (भिन्न)
-
भिन्न
-
240 observations पर 22K+ tokens
-
-
-
Real-time व्यूअर
-
हाँ (port 3113)
-
Cloud dashboard
-
Cloud dashboard
-
Web UI
-
Cloud dashboard
-
Hub web UI
-
नहीं
-
नहीं
-
नहीं
-
नहीं
-
-
-
Self-hosted
-
हाँ (default)
-
Optional
-
Optional
-
हाँ
-
नहीं (केवल-cloud)
-
हाँ (Docker)
-
हाँ
-
हाँ (Oracle DB)
-
हाँ
-
हाँ
-
-
-
-Benchmark नोट: केवल agentmemory का R@5 हमारा अपना measured result है (LongMemEval-S, benchmark/COMPARISON.md से reproducible)। mem0 और Letta के आँकड़े उनके published LoCoMo numbers हैं (एक अलग dataset); MemPalace, supermemory, TencentDB (PersonaMem), और oracleagentmemory के आँकड़े vendor self-reported दावे हैं जिन्हें हमने स्वतंत्र रूप से reproduce नहीं किया है (oracleagentmemory की run ने Oracle AI Database के विरुद्ध GPT-5.5 का उपयोग किया)। केवल ballpark के लिए साथ-साथ दिखाए गए हैं, समान data पर head-to-head तुलना नहीं। Star counts अनुमानित हैं और समय के साथ बदलते रहते हैं।
-
-**नए प्रवेशक** जिन्हें जानना उपयोगी है, [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) में गहराई से compare किए गए:
-
-| System | ⭐ | Angle |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporal knowledge graph; सबसे मज़बूत published temporal-query results (LongMemEval 63.8%), लेकिन graph asynchronously build होता है इसलिए ताज़ा facts पिछड़ सकते हैं |
-| Cognee | 30K | Document-to-knowledge-graph ingestion, केवल-Python, session capture के बजाय structured entity extraction के लिए बना |
+स्रोतों सहित उत्पादों और परिनियोजन विकल्पों की अद्यतन तुलना:
-AgentMemory सत्र capture, खोजने योग्य प्रोजेक्ट इतिहास और लोकल व्यूअर साथ देता है; पहली बार से ही बिना कुंजी के काम कर सकते हैं।
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -780,7 +572,7 @@ Verify with `curl http://localhost:3111/agentmemory/health`. Open http://localho
memory server शुरू करें: `npx -y @agentmemory/agentmemory@latest`
-#### `npx skills add` के माध्यम से Native skills (50+ agents)
+#### `npx skills add` के माध्यम से Native skills
agentmemory Claude-Code-style `/SKILL.md` format में 17 skills ship करता है: 9 invocable action skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) और 8 reference skills जिन्हें agent on demand load करता है (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`)। Reference skills में source से generated data tables होती हैं, इसलिए वे कभी drift नहीं करती। vercel-labs की [`skills`](https://npmjs.com/package/skills) CLI इन्हें calling agent की native skill directory में 50+ agents (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, और अधिक) में auto-install करती है:
@@ -837,7 +629,7 @@ agentmemory entry, `mcpServers` shape का उपयोग करने वा
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` shared customization directory में MCP और capture hooks install करता है। देखें [Antigravity setup और limits](../docs/plugins/antigravity.md)। |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` current IDE versions जैसा ही MCP और hook configuration उपयोग करता है। मौजूदा installations को `--force` से refresh करना चाहिए; देखें [upgrade notes](../docs/plugins/antigravity.md)। |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` user-level config लिखता है। Workspace overrides आपके code के बगल में `.kiro/settings/mcp.json` में जाते हैं। |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` standard `mcpServers` block लिखता है। Warp `.claude/skills/` से skills भी auto-discover करता है; Claude Code plugin install होने के बाद 8 agentmemory skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) Warp की slash-command palette में natively दिखती हैं। |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` standard `mcpServers` block लिखता है। Warp `.claude/skills/` से skills भी auto-discover करता है; Claude Code plugin install होने के बाद 9 agentmemory skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) Warp की slash-command palette में natively दिखती हैं। |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` standard `mcpServers` block लिखता है। VS Code extension users: वही block Cline Settings → MCP Servers → Edit JSON के माध्यम से paste करें। |
| **Continue.dev** | `~/.continue/config.yaml` (preferred) या `config.json` (legacy) | `agentmemory connect continue` कॉन्फ़िगरेशन न होने पर `config.yaml` बनाता है या मौजूद `config.json` अपडेट करता है। पहले से मौजूद YAML के comments और anchors बचाने के लिए उसे नहीं बदलता, बल्कि `mcpServers:` के नीचे पेस्ट करने का block दिखाता है। Continue का `mcpServers` एक array है। |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` `context_servers` (Zed की key, `mcpServers` नहीं) के अंतर्गत लिखता है। Remote MCP servers इसके बजाय `{"url": "..."}` के माध्यम से wired हो सकते हैं। |
@@ -936,7 +728,6 @@ Session 2: "Now add rate limiting"
|---|---|---|
| देखभाल | हाथ से व्यवस्थित निर्देश | इंडेक्स किया गया सत्र इतिहास और मेमोरी |
| खोज | सब कुछ context में load करता है | BM25 + vector + graph (केवल top-K) |
-| Token cost | 240 observations पर 22K+ | ~1,900 tokens (92% कम) |
| Cross-agent | प्रति-agent फाइलें | MCP + REST (कोई भी agent) |
| Coordination | कोई नहीं | Leases, signals, actions, routines |
| Observability | फाइलें मैनुअल पढ़ें | :3113 पर real-time viewer |
@@ -949,7 +740,7 @@ Session 2: "Now add rate limiting"
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1034,7 +825,7 @@ Reciprocal Rank Fusion (RRF, k=60) के साथ fuse होता है औ
जब vector index populated होता है, तो `mem::search` (`memory_recall` के पीछे) hybrid BM25 + vector ranker का उपयोग करता है। Embeddings के बिना यह BM25 का उपयोग करता है। Graph data मौजूद होने पर `smart-search` अतिरिक्त रूप से structural graph matches को भी fuse कर सकता है, keyless mode में भी। Lesson recall हर query पर पूरे corpus को scan करने के बजाय एक dedicated in-memory BM25 index पर चलती है। Superseded memory versions हर recall path से excluded हैं; version chain उनका history रखती है।
-Vectors किसी crash या force-kill से बच जाते हैं। Vector index को कम से कम हर `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 मिनट) पर buckets में save किया जाता है। इसके बीच जोड़ा या हटाया गया हर vector भी state store में एक छोटे pending log में तुरंत लिखा जाता है, और अगला start embedding provider को call किए बिना उसे replay करता है। हर successful save उस log को खाली कर देता है। जिन documents के पास replay के बाद भी कोई vector नहीं है उन्हें background में `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) के batches में तब तक re-embed किया जाता है जब तक कोई बाकी न रहे, और रुका हुआ backfill अगले start पर जारी रहता है। `/agentmemory/status` और viewer pending log का size और backfill state दिखाते हैं। Keyless installs कुछ नहीं लिखते।
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 box से बाहर ही Greek, Cyrillic, Hebrew, Arabic, और accented Latin को tokenize करता है। Chinese / Japanese / Korean memories के लिए, CJK runs को word-level tokens में split करने के लिए optional segmenters install करें (`npm install @node-rs/jieba tiny-segmenter`); उनके बिना, agentmemory soft-fall back होकर whole-run tokenization पर जाता है और stderr पर एक-बार hint print करता है।
@@ -1050,9 +841,9 @@ Normal npm install में optional `@huggingface/transformers` runtime शा
| Provider | Model | Cost | नोट्स |
|---|---|---|---|
-| **Local (अनुशंसित opt-in)** | `all-MiniLM-L6-v2` | Free | पहले model download के बाद on-device, BM25-only पर +8pp recall |
+| **Local (अनुशंसित opt-in)** | `all-MiniLM-L6-v2` | Free | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Free tier | 100+ भाषाएँ, 768/1536/3072 dims (MRL), 2048-token input। `text-embedding-004` को replace करता है ([deprecated, 14 जनवरी 2026 को shutdown](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | उच्चतम quality |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Paid | Code के लिए optimized |
| Cohere | `embed-english-v3.0` | Free trial | General purpose |
| OpenRouter | कोई भी model | भिन्न | Multi-model proxy |
diff --git a/READMEs/README.hu-HU.md b/READMEs/README.hu-HU.md
index fc2ae76681..a5e23b1e78 100644
--- a/READMEs/README.hu-HU.md
+++ b/READMEs/README.hu-HU.md
@@ -66,8 +66,6 @@
-
-> Embedding modell: `all-MiniLM-L6-v2` (lokális, ingyenes, nincs szükség API-kulcsra). Teljes jelentések: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Versenytárs-összehasonlítás: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md), amely az agentmemoryt a mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo rendszerekkel veti össze.
-
-**Lokális reprodukálás:** [`eval/README.md`](../eval/README.md), egy adapter-bővíthető harness a LongMemEval `_s` (nyilvános, 500 kérdéses) + `coding-agent-life-v1` (házon belüli, 15 sessiont tartalmazó korpusz) számára. A grep / vektor / agentmemory adapterek egymás mellett kapnak pontszámot, NDJSON kimenettel, a publikált scorecardok a [`docs/benchmarks/`](../docs/benchmarks/) mappában landolnak.
-
-**Jól illik a [codegraph](https://github.com/colbymchenry/codegraph), az [Understand Anything](https://github.com/Lum1104/Understand-Anything) és a [Graphify](https://github.com/safishamsi/graphify) mellé.** Kódgráf-indexelés, több ágenses build-pipeline-ok, és szélesebb tudásgráfok dokumentumok / PDF-ek / képek / videók fölött. Az agentmemory a munkára emlékszik; ez a három projekt a kontextusréteg többi részét kelti életre. Receptek + kérdés-útválasztási táblázat: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Beépített (CLAUDE.md)
-
-
-
Típus
-
Memóriamotor + MCP szerver
-
Memóriaréteg API
-
Teljes ágens-futtatókörnyezet
-
Személyes AI
-
Memória API + app
-
Csapatmemória-hub (LLM proxy)
-
Vektormemória (OSS)
-
Memóriamotor (Oracle DB)
-
Memóriarendszer
-
Statikus fájl
-
-
-
Visszakeresési R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Önbevallás alapján
-
PersonaMem 76% (önbevallás alapján)
-
~96.6% (önbevallás alapján)
-
94.4% (önbevallás alapján)
-
N/A
-
N/A (grep)
-
-
-
Automatikus rögzítés
-
12 hook (nulla manuális munka)
-
Manuális add() hívások
-
Az ágens saját szerkesztései
-
Manuális
-
API-oldali kinyerés
-
Proxy-elfogás (base-URL csere)
-
Manuális
-
API kinyerés
-
Manuális
-
Manuális szerkesztés
-
-
-
Keresés
-
BM25 + Vektor + Gráf (RRF fúzió)
-
Vektor + Gráf
-
Vektor (archivális)
-
Szemantikus
-
Vektor + RAG
-
4 eszköztípus (Chat / Skill / Wiki / CodeGraph)
-
Csak vektor
-
Vektor + szemantikus
-
Lecsengés-súlyozott
-
Mindent betölt a kontextusba
-
-
-
Több ágens
-
MCP + REST + lease-ek + szignálok
-
API (nincs koordináció)
-
Csak a Letta futtatókörnyezeten belül
-
Nem
-
Nem
-
Csapatszerepek + megosztott eszközök
-
Nem
-
Csak hatókörön belül
-
Több ágens közt megosztva
-
Ágensenkénti fájlok
-
-
-
Keretrendszer-függés
-
Nincs (bármely MCP kliens)
-
Nincs
-
Magas (a Lettát kell használni)
-
Önálló
-
Nincs
-
A proxy minden modellhívás elé kerül
-
Nincs
-
Oracle Database
-
Nincs
-
Ágensenkénti formátum
-
-
-
Külső függőségek
-
Nem szükséges külön adatbázis-szolgáltatás
-
Qdrant / pgvector
-
Postgres + vektor DB
-
Több
-
Felügyelt felhő
-
Docker-stack (Core + Hub + Proxy)
-
Vektortár
-
Oracle AI Database
-
Nincs
-
Nincs
-
-
-
Memória életciklus
-
4 szintű konszolidáció + lecsengés + automatikus elfelejtés
-
Passzív kinyerés
-
Az ágens kezeli
-
Manuális
-
Automatikus elfelejtés
-
Manuális átnézés; az automatikus útválasztás folyamatban
-
Nincs
-
Nincs megadva
-
Lecsengés + konszolidáció
-
Manuális ritkítás
-
-
-
Token-hatékonyság
-
~1,900 token/session ($10/év)
-
Az integrációtól függ
-
A core memória a kontextusban van
-
Változó
-
Felhő alapú díjszabás
-
Nincs megadva
-
Nincs token-budget
-
LLM-alapú (változó)
-
Változó
-
22K+ token 240 megfigyelésnél
-
-
-
Valós idejű megjelenítő
-
Igen (3113-as port)
-
Felhő alapú dashboard
-
Felhő alapú dashboard
-
Webes felhasználói felület
-
Felhő alapú dashboard
-
Hub webes felhasználói felület
-
Nincs
-
Nincs
-
Nincs
-
Nincs
-
-
-
Önállóan üzemeltethető
-
Igen (alapértelmezett)
-
Opcionális
-
Opcionális
-
Igen
-
Nem (csak felhő)
-
Igen (Docker)
-
Igen
-
Igen (Oracle DB)
-
Igen
-
Igen
-
-
-
-Megjegyzés a benchmarkhoz: csak az agentmemory R@5 értéke a saját, mért eredményünk (LongMemEval-S, reprodukálható a benchmark/COMPARISON.md alapján). A mem0 és a Letta számai a saját publikált LoCoMo-adataik (egy másik adathalmaz); a MemPalace, a supermemory, a TencentDB (PersonaMem) és az oracleagentmemory számai a gyártók önbevallás alapú állításai, amelyeket nem reprodukáltunk függetlenül (az oracleagentmemory futása GPT-5.5-öt használt egy Oracle AI Database ellen). Csak tájékoztató jellegű egymás mellett közlésre szánva, nem azonos adatokon futó közvetlen összehasonlításra. A csillagszámok közelítők, és idővel változnak.
-
-**Újabb belépők**, amelyeket érdemes ismerni, részletesen összehasonlítva a [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) fájlban:
-
-| Rendszer | ⭐ | Megközelítés |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporális tudásgráf; a legerősebb publikált időbeli lekérdezési eredmények (LongMemEval 63.8%), de a gráf aszinkron módon épül, így a friss tények lemaradhatnak |
-| Cognee | 30K | Dokumentum-tudásgráf feldolgozás, csak Python, strukturált entitáskinyerésre épül, nem session-rögzítésre |
+A termékek és telepítési lehetőségek frissített összehasonlítása, forrásokkal:
-Az AgentMemory egyesíti a munkamenetek rögzítését, a kereshető projektelőzményeket és a helyi nézőfelületet, az első indítástól kulcs nélküli működéssel.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -784,7 +576,7 @@ Teljes útmutató: [`integrations/hermes/`](../integrations/hermes/)
Indítsd el a memóriaszervert: `npx -y @agentmemory/agentmemory@latest`
-#### Natív skillek az `npx skills add` révén (50+ ágens)
+#### Natív skillek az `npx skills add` révén
Az agentmemory 17 skillt szállít a Claude-Code-stílusú `/SKILL.md` formátumban: 9 meghívható akció-skill (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) és 8 referencia-skill, amelyeket az ágens igény szerint tölt be (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). A referencia-skillek forrásból generált adattáblákat hordoznak, így soha nem térnek el. A [`skills`](https://npmjs.com/package/skills) CLI, a vercel-labs eszköze, automatikusan telepíti ezeket a hívó ágens natív skill-mappájába, 50+ ágens között (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, és még több):
@@ -842,7 +634,7 @@ Az agentmemory bejegyzés **ugyanaz az MCP szerver blokk** minden olyan hoszton,
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | Az `agentmemory connect antigravity --with-hooks` telepíti az MCP-t és a rögzítő hookokat a megosztott testreszabási mappába. Lásd az [Antigravity beállítása és korlátai](../docs/plugins/antigravity.md) útmutatót. |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | Az `agentmemory connect antigravity-cli --with-hooks` ugyanazt az MCP- és hook-konfigurációt használja, mint a jelenlegi IDE-verziók. A meglévő telepítéseknek frissülniük kell a `--force`-szal; lásd a [frissítési megjegyzéseket](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | Az `agentmemory connect kiro` beírja a felhasználói szintű konfigurációt. A workspace-szintű felülírások a `.kiro/settings/mcp.json`-ba kerülnek, a kódod mellé. |
-| **Warp** | `~/.warp/.mcp.json` | Az `agentmemory connect warp` beírja a standard `mcpServers` blokkot. A Warp a skilleket is automatikusan felismeri a `.claude/skills/`-ből; a Claude Code plugin telepítése után a 8 agentmemory skill (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) natívan megjelenik a Warp slash-parancs palettáján. |
+| **Warp** | `~/.warp/.mcp.json` | Az `agentmemory connect warp` beírja a standard `mcpServers` blokkot. A Warp a skilleket is automatikusan felismeri a `.claude/skills/`-ből; a Claude Code plugin telepítése után a 9 agentmemory skill (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) natívan megjelenik a Warp slash-parancs palettáján. |
| **Cline (CLI)** | `~/.cline/mcp.json` | Az `agentmemory connect cline` beírja a standard `mcpServers` blokkot. VS Code extension felhasználóknak: illeszd be ugyanazt a blokkot a Cline Settings → MCP Servers → Edit JSON útján. |
| **Continue.dev** | `~/.continue/config.yaml` (előnyben részesített) vagy `config.json` (örökölt) | Az `agentmemory connect continue` létrehozza a hiányzó `config.yaml` fájlt, vagy módosítja a meglévő `config.json` fájlt. Meglévő YAML esetén kiírja az `mcpServers:` alá illeszthető blokkot, megőrizve a megjegyzéseket és horgonyokat. A Continue tömböt használ, nem objektumot. |
| **Zed** | `~/.config/zed/settings.json` | Az `agentmemory connect zed` a `context_servers` alá ír (a Zed saját kulcsa, NEM a `mcpServers`). A távoli MCP szerverek egy `{"url": "..."}`-vel is bekötődhetnek. |
@@ -925,7 +717,6 @@ Minden AI kódolóágens beépített memóriával érkezik: a Claude Code-nak `M
|---|---|---|
| Karbantartás | Kézzel gondozott instrukciók | Indexelt munkamenet-előzmények és emlékek |
| Keresés | Mindent betölt a kontextusba | BM25 + vektor + gráf (csak top-K) |
-| Token-költség | 22K+ 240 megfigyelésnél | ~1,900 token (92%-kal kevesebb) |
| Ágensek közötti | Ágensenkénti fájlok | MCP + REST (bármely ágens) |
| Koordináció | Nincs | Lease-ek, szignálok, akciók, rutinok |
| Megfigyelhetőség | Fájlok manuális olvasása | Valós idejű megjelenítő a :3113-on |
@@ -938,7 +729,7 @@ Minden AI kódolóágens beépített memóriával érkezik: a Claude Code-nak `M
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1023,7 +814,7 @@ Reciprocal Rank Fusionnal (RRF, k=60) fuzionálva, és sessionönkénti diverzif
Ha egy vektorindex fel van töltve, a `mem::search` (a `memory_recall` mögött) a hibrid BM25 + vektor rangsorolót használja. Embeddingek nélkül BM25-öt használ. A `smart-search` emellett strukturális gráfegyezéseket is fuzionálhat, ha létezik gráfadat, kulcs nélküli módban is. A tanulság-felidézés egy dedikált, memóriában tartott BM25-indexen fut, nem a teljes korpusz lekérdezésenkénti átnézésével. A felülírt memória-verziók minden felidézési útvonalból ki vannak zárva; a verziólánc megtartja az előzményüket.
-A vektorok túlélnek egy összeomlást vagy egy force-killt. A vektorindex legfeljebb minden `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS`-ben (10 perc) kerül mentésre, csomagokban. Minden közben hozzáadott vagy eltávolított vektor azonnal felíródik egy kis, függőben lévő naplóba is az állapottárban, és a következő indítás lejátssza ezt az embedding-szolgáltató hívása nélkül. Minden sikeres mentés kiüríti a naplót. Azok a dokumentumok, amelyeknek a lejátszás után sincs vektoruk, a háttérben, `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) méretű batch-ekben kapnak újra embeddinget, amíg nem marad ilyen, és egy leállított backfill a következő indításkor folytatódik. A `/agentmemory/status` és a megjelenítő megmutatja a függőben lévő napló méretét és a backfill állapotát. A kulcs nélküli telepítések semmit nem írnak.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
A BM25 alapból tokenizálja a görögöt, a cirillt, a hébert, az arabot és az ékezetes latint. Kínai / japán / koreai memóriákhoz telepítsd az opcionális szegmentálókat (`npm install @node-rs/jieba tiny-segmenter`), hogy szóméretű tokenekre bontsa a CJK-futásokat; nélkülük az agentmemory finoman visszaesik a teljes futás tokenizálására, és egyszeri jelzést ír a stderr-re.
@@ -1039,9 +830,9 @@ A normál npm-telepítés tartalmazza az opcionális `@huggingface/transformers`
| Szolgáltató | Modell | Költség | Megjegyzések |
|---|---|---|---|
-| **Lokális (ajánlott opt-in)** | `all-MiniLM-L6-v2` | Ingyenes | Eszközön fut az első modell-letöltés után, +8pp felidézés a csak-BM25-höz képest |
+| **Lokális (ajánlott opt-in)** | `all-MiniLM-L6-v2` | Ingyenes | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Ingyenes szint | 100+ nyelv, 768/1536/3072 dimenzió (MRL), 2048 tokenes bemenet. Felváltja a `text-embedding-004`-et ([elavult, leállítva 2026. jan. 14-én](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Legjobb minőség |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Fizetős | Kódra optimalizálva |
| Cohere | `embed-english-v3.0` | Ingyenes próba | Általános célú |
| OpenRouter | Bármely modell | Változó | Több modellt lefedő proxy |
diff --git a/READMEs/README.id-ID.md b/READMEs/README.id-ID.md
index 1de7ed8084..06a24851a6 100644
--- a/READMEs/README.id-ID.md
+++ b/READMEs/README.id-ID.md
@@ -66,8 +66,6 @@
-
-> Model embedding: `all-MiniLM-L6-v2` (lokal, gratis, tanpa API key). Laporan lengkap: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Perbandingan kompetitor: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) mencakup agentmemory vs mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Reproduksi secara lokal:** [`eval/README.md`](../eval/README.md), harness yang adapternya bisa dipasang-cabut untuk LongMemEval `_s` (500-Q publik) + `coding-agent-life-v1` (korpus in-house 15 sesi). Adapter grep / vector / agentmemory dinilai berdampingan, output NDJSON, scorecard yang dipublikasikan mendarat di [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Berpasangan dengan [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything), dan [Graphify](https://github.com/safishamsi/graphify).** Code-graph indexing, pipeline build multi-agen, dan knowledge graph yang lebih luas lintas dokumen / PDF / gambar / video. agentmemory mengingat pekerjaannya; ketiga proyek itu menghidupkan sisa layer konteksnya. Resep + tabel perutean pertanyaan: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Bawaan (CLAUDE.md)
-
-
-
Tipe
-
Memory engine + server MCP
-
API memory layer
-
Runtime agen lengkap
-
AI pribadi
-
API memori + app
-
Hub memori tim (proxy LLM)
-
Vector memory (OSS)
-
Memory engine (Oracle DB)
-
Sistem memori
-
File statis
-
-
-
Retrieval R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Self-reported
-
PersonaMem 76% (self-reported)
-
~96.6% (self-reported)
-
94.4% (self-reported)
-
N/A
-
N/A (grep)
-
-
-
Auto-capture
-
12 hook (tanpa effort manual)
-
Panggilan add() manual
-
Self-edit oleh agen
-
Manual
-
Ekstraksi di sisi API
-
Intersepsi proxy (swap base-URL)
-
Manual
-
Ekstraksi API
-
Manual
-
Edit manual
-
-
-
Search
-
BM25 + Vector + Graph (fusi RRF)
-
Vector + Graph
-
Vector (archival)
-
Semantik
-
Vector + RAG
-
4 tipe aset (Chat / Skill / Wiki / CodeGraph)
-
Vector-only
-
Vector + semantik
-
Decay-weighted
-
Memuat semuanya ke konteks
-
-
-
Multi-agen
-
MCP + REST + lease + signal
-
API (tanpa koordinasi)
-
Hanya dalam runtime Letta
-
Tidak
-
Tidak
-
Role tim + aset bersama
-
Tidak
-
Hanya terbatas (scoped)
-
Multi-agen bersama
-
File per agen
-
-
-
Framework lock-in
-
Tidak ada (klien MCP apa pun)
-
Tidak ada
-
Tinggi (harus pakai Letta)
-
Standalone
-
Tidak ada
-
Proxy mengarahkan setiap panggilan model
-
Tidak ada
-
Oracle Database
-
Tidak ada
-
Format per agen
-
-
-
Dependensi eksternal
-
Tidak memerlukan layanan database terpisah
-
Qdrant / pgvector
-
Postgres + vector DB
-
Multipel
-
Cloud terkelola
-
Stack Docker (Core + Hub + Proxy)
-
Vector store
-
Oracle AI Database
-
Tidak ada
-
Tidak ada
-
-
-
Siklus hidup memori
-
Konsolidasi 4-tier + decay + auto-forget
-
Ekstraksi pasif
-
Dikelola agen
-
Manual
-
Auto-forget
-
Review manual; auto-routing sedang dikerjakan
-
Tidak ada
-
Tidak disebutkan
-
Decay + konsolidasi
-
Pruning manual
-
-
-
Efisiensi token
-
~1.900 token/sesi ($10/thn)
-
Bervariasi menurut integrasi
-
Core memory di konteks
-
Bervariasi
-
Harga cloud
-
Tidak disebutkan
-
Tanpa budget token
-
Didukung LLM (bervariasi)
-
Bervariasi
-
22K+ token pada 240 observasi
-
-
-
Viewer real-time
-
Ya (port 3113)
-
Dashboard cloud
-
Dashboard cloud
-
Web UI
-
Dashboard cloud
-
Web UI Hub
-
Tidak
-
Tidak
-
Tidak
-
Tidak
-
-
-
Self-hosted
-
Ya (default)
-
Opsional
-
Opsional
-
Ya
-
Tidak (cloud-only)
-
Ya (Docker)
-
Ya
-
Ya (Oracle DB)
-
Ya
-
Ya
-
-
-
-Catatan benchmark: hanya R@5 agentmemory yang merupakan hasil pengukuran kami sendiri (LongMemEval-S, bisa direproduksi dari benchmark/COMPARISON.md). Angka mem0 dan Letta adalah angka LoCoMo yang mereka publikasikan sendiri (dataset yang berbeda); angka MemPalace, supermemory, TencentDB (PersonaMem), dan oracleagentmemory adalah klaim self-reported vendor yang belum kami reproduksi secara independen (run oracleagentmemory memakai GPT-5.5 terhadap Oracle AI Database). Ditampilkan berdampingan hanya sebagai perkiraan kasar, bukan head-to-head pada data yang identik. Jumlah star bersifat perkiraan dan bisa berubah seiring waktu.
-
-**Pemain baru** yang layak diketahui, dibahas lebih dalam di [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Sistem | ⭐ | Sudut pandang |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Knowledge graph temporal; hasil query temporal terpublikasi terkuat (LongMemEval 63.8%), tapi graph dibangun secara asinkron sehingga fakta baru bisa tertinggal |
-| Cognee | 30K | Ingestion dokumen-ke-knowledge-graph, Python-only, dibangun untuk ekstraksi entitas terstruktur bukan capture sesi |
+Perbandingan produk dan opsi penerapan terkini, beserta sumbernya:
-AgentMemory memadukan capture sesi, riwayat proyek yang bisa dicari, dan viewer lokal, dengan mode tanpa kunci sejak awal.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -780,7 +572,7 @@ Panduan lengkap: [`integrations/hermes/`](../integrations/hermes/)
Jalankan server memori: `npx -y @agentmemory/agentmemory@latest`
-#### Native skill lewat `npx skills add` (50+ agen)
+#### Native skill lewat `npx skills add`
agentmemory menyediakan 17 skill dalam format `/SKILL.md` ala Claude Code: 9 skill aksi yang bisa dipanggil (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) dan 8 skill referensi yang dimuat agen sesuai kebutuhan (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Skill referensi membawa tabel data yang dihasilkan dari source, sehingga tidak pernah drift. CLI [`skills`](https://npmjs.com/package/skills) oleh vercel-labs memasangnya secara otomatis ke direktori skill native agen yang memanggilnya, di 50+ agen (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, dan lainnya):
@@ -837,7 +629,7 @@ Entri agentmemory adalah **blok server MCP yang sama** di setiap host yang memak
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` memasang MCP dan hook capture di direktori kustomisasi bersama. Lihat [pengaturan dan batasan Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` memakai konfigurasi MCP dan hook yang sama dengan versi IDE saat ini. Instalasi yang sudah ada sebaiknya di-refresh dengan `--force`; lihat [catatan upgrade](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` menulis konfigurasi tingkat pengguna. Override workspace masuk di `.kiro/settings/mcp.json` di sebelah kode Anda. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` menulis blok `mcpServers` standar. Warp juga auto-discover skill dari `.claude/skills/`; setelah plugin Claude Code terpasang, 8 skill agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) muncul secara native di palet slash-command Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` menulis blok `mcpServers` standar. Warp juga auto-discover skill dari `.claude/skills/`; setelah plugin Claude Code terpasang, 9 skill agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) muncul secara native di palet slash-command Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` menulis blok `mcpServers` standar. Pengguna ekstensi VS Code: tempel blok yang sama lewat Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (disarankan) atau `config.json` (legacy) | `agentmemory connect continue` membuat `config.yaml` jika belum ada konfigurasi atau memperbarui `config.json` yang ada. YAML yang sudah ada dipertahankan agar komentar dan anchor tetap utuh; perintah mencetak blok untuk ditempel di bawah `mcpServers:`. Continue memakai bentuk array untuk `mcpServers`. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` menulis di bawah `context_servers` (key milik Zed, BUKAN `mcpServers`). Server MCP remote bisa disambungkan lewat `{"url": "..."}` sebagai gantinya. |
@@ -934,7 +726,6 @@ Setiap agen coding AI dilengkapi memori bawaan: Claude Code punya `MEMORY.md`, C
|---|---|---|
| Pemeliharaan | Instruksi yang dikelola manual | Riwayat sesi dan memori terindeks |
| Search | Memuat semuanya ke konteks | BM25 + vector + graph (hanya top-K) |
-| Biaya token | 22K+ pada 240 observasi | ~1.900 token (92% lebih sedikit) |
| Lintas agen | File per agen | MCP + REST (agen apa pun) |
| Koordinasi | Tidak ada | Lease, signal, action, routine |
| Observability | Baca file secara manual | Viewer real-time di :3113 |
@@ -947,7 +738,7 @@ Setiap agen coding AI dilengkapi memori bawaan: Claude Code punya `MEMORY.md`, C
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1032,7 +823,7 @@ Difusikan dengan Reciprocal Rank Fusion (RRF, k=60) dan didiversifikasi per sesi
Saat index vector terisi, `mem::search` (di belakang `memory_recall`) memakai ranker hybrid BM25 + vector. Tanpa embedding ia memakai BM25. `smart-search` juga bisa menggabungkan kecocokan graph struktural saat data graph ada, termasuk di mode keyless. Recall lesson berjalan di index BM25 in-memory khusus bukan memindai seluruh korpus per query. Versi memori yang disupersede dikecualikan dari setiap jalur recall; rantai versi menyimpan histori mereka.
-Vector tetap bertahan dari crash atau force-kill. Index vector disimpan dalam bucket paling lambat setiap `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 menit). Setiap vector yang ditambahkan atau dihapus di antaranya juga langsung ditulis ke log pending kecil di state store, dan start berikutnya me-replay-nya tanpa memanggil penyedia embedding. Setiap save yang berhasil mengosongkan log itu. Dokumen yang masih belum punya vector setelah replay di-embed ulang di background dalam batch `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) sampai tidak ada yang tersisa, dan backfill yang terhenti dilanjutkan pada start berikutnya. `/agentmemory/status` dan viewer menunjukkan ukuran log pending dan state backfill. Instalasi keyless tidak menulis apa pun.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 men-tokenisasi Yunani, Sirilik, Ibrani, Arab, dan Latin beraksen secara bawaan. Untuk memori Tionghoa / Jepang / Korea, pasang segmenter opsional (`npm install @node-rs/jieba tiny-segmenter`) untuk membelah runtun CJK menjadi token level-kata; tanpa itu, agentmemory soft-fallback ke tokenisasi seluruh runtun dan mencetak hint satu-kali di stderr.
@@ -1048,9 +839,9 @@ Instalasi npm normal menyertakan runtime opsional `@huggingface/transformers`. P
| Penyedia | Model | Biaya | Catatan |
|---|---|---|---|
-| **Lokal (opt-in yang disarankan)** | `all-MiniLM-L6-v2` | Gratis | Di perangkat sendiri setelah unduhan model pertama, +8pp recall dibanding BM25-only |
+| **Lokal (opt-in yang disarankan)** | `all-MiniLM-L6-v2` | Gratis | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Free tier | 100+ bahasa, 768/1536/3072 dimensi (MRL), input 2048-token. Menggantikan `text-embedding-004` ([deprecated, shutdown 14 Jan 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Kualitas tertinggi |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Berbayar | Dioptimalkan untuk code |
| Cohere | `embed-english-v3.0` | Free trial | Tujuan umum |
| OpenRouter | Model apa pun | Bervariasi | Proxy multi-model |
diff --git a/READMEs/README.it-IT.md b/READMEs/README.it-IT.md
index 3e3fa649d4..871d409347 100644
--- a/READMEs/README.it-IT.md
+++ b/READMEs/README.it-IT.md
@@ -66,8 +66,6 @@
-
-### Precisione del retrieval
-
-**coding-agent-life-v1** (corpus interno, riproducibile in sandbox)
-
-| Adapter | P@5 | R@5 | Tasso di successo top-5 | latenza p50 |
-|---|---|---|---|---|
-| **agentmemory ibrido** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep di base | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-Tasso di successo top-5 del 100% al **limite matematico di P@5** per questo corpus (0.240, vedi scorecard). L'ibrido recupera ogni sessione gold; grep ne manca 1 su 2 nella query temporale multi-sessione. Il guadagno è su **recall + temporale**, non sulla precisione aggregata. Questo benchmark è piccolo e povero di dati gold; il più ampio LongMemEval-S sotto differenzia meglio. Analisi completa per tipo + nota di correzione: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 domande)
+Risultati attuali del recupero, definizioni delle metriche e procedura di riproduzione:
-| Sistema | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| Fallback solo BM25 | 86.2% | 94.6% | 71.5% |
-
-
-
-> Modello di embedding: `all-MiniLM-L6-v2` (locale, gratuito, senza chiave API). Report completi: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Confronto con i concorrenti: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) che copre agentmemory vs mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Riproducibile in locale:** [`eval/README.md`](../eval/README.md), un harness con adapter collegabili per LongMemEval `_s` (500 domande pubbliche) + `coding-agent-life-v1` (corpus interno di 15 sessioni). Gli adapter grep / vettoriale / agentmemory ottengono un punteggio fianco a fianco, output NDJSON, le scorecard pubblicate finiscono in [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Si combina con [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) e [Graphify](https://github.com/safishamsi/graphify).** Indicizzazione code-graph, pipeline di build multi-agente e grafi di conoscenza più ampi su documenti / PDF / immagini / video. agentmemory ricorda il lavoro; questi tre progetti illuminano il resto dello strato di contesto. Ricette + tabella di instradamento delle domande: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Integrata (CLAUDE.md)
-
-
-
Tipo
-
Motore di memoria + server MCP
-
API livello memoria
-
Runtime agente completo
-
IA personale
-
API memoria + app
-
Hub memoria di team (proxy LLM)
-
Memoria vettoriale (OSS)
-
Motore di memoria (Oracle DB)
-
Sistema di memoria
-
File statico
-
-
-
Retrieval R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/D
-
Auto-dichiarato
-
PersonaMem 76% (auto-dichiarato)
-
~96.6% (auto-dichiarato)
-
94.4% (auto-dichiarato)
-
N/D
-
N/D (grep)
-
-
-
Acquisizione automatica
-
12 hook (zero sforzo manuale)
-
Chiamate manuali add()
-
Auto-modifiche dell'agente
-
Manuale
-
Estrazione lato API
-
Intercettazione proxy (sostituzione base-URL)
-
Manuale
-
Estrazione API
-
Manuale
-
Modifica manuale
-
-
-
Ricerca
-
BM25 + vettoriale + grafo (fusione RRF)
-
Vettoriale + grafo
-
Vettoriale (archivio)
-
Semantica
-
Vettoriale + RAG
-
4 tipi di asset (Chat / Skill / Wiki / CodeGraph)
-
Solo vettoriale
-
Vettoriale + semantica
-
Pesata per decadimento
-
Carica tutto nel contesto
-
-
-
Multi-agente
-
MCP + REST + lease + segnali
-
API (nessun coordinamento)
-
Solo all'interno del runtime Letta
-
No
-
No
-
Ruoli di team + asset condivisi
-
No
-
Solo con ambito limitato
-
Condiviso multi-agente
-
File per agente
-
-
-
Vincolo al framework
-
Nessuno (qualsiasi client MCP)
-
Nessuno
-
Alto (deve usare Letta)
-
Autonomo
-
Nessuno
-
Il proxy intercetta ogni chiamata al modello
-
Nessuno
-
Oracle Database
-
Nessuno
-
Formato per agente
-
-
-
Dipendenze esterne
-
Nessun servizio database separato richiesto
-
Qdrant / pgvector
-
Postgres + DB vettoriale
-
Multiple
-
Cloud gestito
-
Stack Docker (Core + Hub + Proxy)
-
Vector store
-
Oracle AI Database
-
Nessuna
-
Nessuna
-
-
-
Ciclo di vita della memoria
-
Consolidamento a 4 livelli + decadimento + auto-oblio
-
Estrazione passiva
-
Gestito dall'agente
-
Manuale
-
Auto-oblio
-
Revisione manuale; instradamento automatico in corso
-
Nessuno
-
Non dichiarato
-
Decadimento + consolidamento
-
Pulizia manuale
-
-
-
Efficienza dei token
-
~1,900 token/sessione ($10/anno)
-
Varia per integrazione
-
Memoria core nel contesto
-
Varia
-
Prezzo cloud
-
Non dichiarato
-
Nessun budget di token
-
Basato su LLM (varia)
-
Varia
-
22K+ token a 240 osservazioni
-
-
-
Viewer in tempo reale
-
Sì (porta 3113)
-
Dashboard cloud
-
Dashboard cloud
-
UI web
-
Dashboard cloud
-
UI web Hub
-
No
-
No
-
No
-
No
-
-
-
Self-hosted
-
Sì (predefinito)
-
Opzionale
-
Opzionale
-
Sì
-
No (solo cloud)
-
Sì (Docker)
-
Sì
-
Sì (Oracle DB)
-
Sì
-
Sì
-
-
-
-Nota sul benchmark: solo l'R@5 di agentmemory è un risultato misurato da noi (LongMemEval-S, riproducibile da benchmark/COMPARISON.md). I valori di mem0 e Letta sono i loro numeri pubblicati su LoCoMo (un dataset diverso); i valori di MemPalace, supermemory, TencentDB (PersonaMem) e oracleagentmemory sono dichiarazioni auto-riportate dai rispettivi fornitori che non abbiamo riprodotto in modo indipendente (l'esecuzione di oracleagentmemory ha usato GPT-5.5 contro un Oracle AI Database). Mostrati fianco a fianco solo per un confronto approssimativo, non un testa a testa su dati identici. I conteggi delle stelle sono approssimativi e cambiano nel tempo.
-
-**Nuovi arrivati** da conoscere, confrontati in profondità in [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Sistema | ⭐ | Approccio |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Grafo di conoscenza temporale; i risultati pubblicati più solidi sulle query temporali (LongMemEval 63.8%), ma il grafo si costruisce in modo asincrono quindi i fatti recenti possono essere in ritardo |
-| Cognee | 30K | Ingestione da documento a grafo di conoscenza, solo Python, costruito per l'estrazione strutturata di entità più che per l'acquisizione di sessioni |
+Confronto aggiornato dei prodotti e delle opzioni di distribuzione, con fonti:
-AgentMemory combina acquisizione dagli agenti di programmazione, cronologia del progetto ricercabile e visualizzatore locale. La modalità senza chiave è disponibile dal primo avvio.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -779,7 +571,7 @@ Guida completa: [`integrations/hermes/`](../integrations/hermes/)
Avvia il server di memoria: `npx -y @agentmemory/agentmemory@latest`
-#### Skill native tramite `npx skills add` (50+ agenti)
+#### Skill native tramite `npx skills add`
agentmemory distribuisce 17 skill nel formato `/SKILL.md` in stile Claude Code: 9 skill d'azione invocabili (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) e 8 skill di riferimento che l'agente carica su richiesta (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Le skill di riferimento portano tabelle di dati generate dalla fonte, quindi non diventano mai obsolete. La CLI [`skills`](https://npmjs.com/package/skills) di vercel-labs le installa automaticamente nella directory delle skill native dell'agente chiamante su 50+ agenti (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, e altri):
@@ -836,7 +628,7 @@ La voce agentmemory è lo **stesso blocco server MCP** su ogni host che usa la f
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` installa l'MCP e gli hook di acquisizione nella directory di personalizzazione condivisa. Consulta [configurazione e limiti di Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` usa la stessa configurazione MCP e hook delle versioni IDE attuali. Le installazioni esistenti dovrebbero aggiornare con `--force`; consulta le [note di aggiornamento](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` scrive la configurazione a livello utente. Gli override per workspace vanno in `.kiro/settings/mcp.json` accanto al tuo codice. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` scrive il blocco standard `mcpServers`. Warp rileva automaticamente anche le skill da `.claude/skills/`; una volta installato il plugin Claude Code, le 8 skill di agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) appaiono nativamente nella palette slash-command di Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` scrive il blocco standard `mcpServers`. Warp rileva automaticamente anche le skill da `.claude/skills/`; una volta installato il plugin Claude Code, le 9 skill di agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) appaiono nativamente nella palette slash-command di Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` scrive il blocco standard `mcpServers`. Utenti dell'estensione VS Code: incolla lo stesso blocco tramite Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (preferito) o `config.json` (legacy) | `agentmemory connect continue` crea `config.yaml` se manca o modifica il `config.json` esistente. Se il file YAML esiste, mostra il blocco da aggiungere a `mcpServers:` per conservare commenti e ancore. Continue usa un array, non un oggetto. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` scrive sotto `context_servers` (la chiave di Zed, NON `mcpServers`). I server MCP remoti possono essere collegati tramite `{"url": "..."}` invece. |
@@ -933,7 +725,6 @@ Ogni agente di coding AI ha una memoria integrata: Claude Code ha `MEMORY.md`, C
|---|---|---|
| Manutenzione | Istruzioni curate manualmente | Cronologia delle sessioni e ricordi indicizzati |
| Ricerca | Carica tutto nel contesto | BM25 + vettoriale + grafo (solo top-K) |
-| Costo in token | 22K+ a 240 osservazioni | ~1,900 token (92% in meno) |
| Cross-agente | File per agente | MCP + REST (qualsiasi agente) |
| Coordinamento | Nessuno | Lease, segnali, azioni, routine |
| Osservabilità | Lettura manuale dei file | Viewer in tempo reale su :3113 |
@@ -946,7 +737,7 @@ Ogni agente di coding AI ha una memoria integrata: Claude Code ha `MEMORY.md`, C
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1031,7 +822,7 @@ Fuso con Reciprocal Rank Fusion (RRF, k=60) e diversificato per sessione (massim
Quando un indice vettoriale è popolato, `mem::search` (dietro `memory_recall`) usa il ranker ibrido BM25 + vettoriale. Senza embedding usa BM25. `smart-search` può inoltre fondere corrispondenze strutturali del grafo quando i dati del grafo esistono, anche in modalità keyless. Il recall delle lezioni funziona su un indice BM25 dedicato in memoria invece di scansionare l'intero corpus per ogni query. Le versioni superate della memoria sono escluse da ogni percorso di recall; la catena delle versioni conserva la loro storia.
-I vettori sopravvivono a un crash o a un force-kill. L'indice vettoriale viene salvato a blocchi almeno ogni `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minuti). Ogni vettore aggiunto o rimosso nel frattempo viene anche scritto immediatamente in un piccolo log in sospeso nello state store, e il prossimo avvio lo riapplica senza chiamare il provider di embedding. Ogni salvataggio riuscito svuota il log. I documenti che restano senza vettore dopo la riapplicazione vengono ri-embeddati in background a lotti di `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) finché non ne resta nessuno, e un backfill interrotto riprende al prossimo avvio. `/agentmemory/status` e il viewer mostrano la dimensione del log in sospeso e lo stato del backfill. Le installazioni keyless non scrivono nulla.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokenizza nativamente greco, cirillico, ebraico, arabo e latino accentato. Per le memorie in cinese / giapponese / coreano, installa i segmentatori opzionali (`npm install @node-rs/jieba tiny-segmenter`) per dividere le sequenze CJK in token a livello di parola; senza di essi, agentmemory ricade in modo graduale sulla tokenizzazione dell'intera sequenza e stampa un suggerimento una tantum su stderr.
@@ -1047,9 +838,9 @@ L'installazione npm normale include il runtime opzionale `@huggingface/transform
| Provider | Modello | Costo | Note |
|---|---|---|---|
-| **Locale (consigliato come opt-in)** | `all-MiniLM-L6-v2` | Gratuito | On-device dopo il primo download del modello, +8pp di recall rispetto a solo BM25 |
+| **Locale (consigliato come opt-in)** | `all-MiniLM-L6-v2` | Gratuito | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Livello gratuito | 100+ lingue, 768/1536/3072 dimensioni (MRL), input di 2048 token. Sostituisce `text-embedding-004` ([deprecato, spegnimento il 14 gennaio 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Qualità più alta |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | A pagamento | Ottimizzato per il codice |
| Cohere | `embed-english-v3.0` | Prova gratuita | Uso generale |
| OpenRouter | Qualsiasi modello | Variabile | Proxy multi-modello |
diff --git a/READMEs/README.ja-JP.md b/READMEs/README.ja-JP.md
index 12b8352cf2..80f4badaf8 100644
--- a/READMEs/README.ja-JP.md
+++ b/READMEs/README.ja-JP.md
@@ -66,8 +66,6 @@
-
-> 임베딩 모델: `all-MiniLM-L6-v2` (로컬, 무료, API 키 불필요). 전체 보고서: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). 경쟁 제품 비교: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) — agentmemory 대 mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**로컬 재현 방법:** [`eval/README.md`](../eval/README.md), LongMemEval `_s`(공개 500-Q)와 `coding-agent-life-v1`(자체 15-세션 코퍼스)을 위한 어댑터 플러그형 하니스입니다. grep / vector / agentmemory 어댑터를 나란히 평가하고, NDJSON으로 출력하며, 게시된 스코어카드는 [`docs/benchmarks/`](../docs/benchmarks/)에 보관됩니다.
-
-**다음과 함께 사용하기 좋습니다: [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything), [Graphify](https://github.com/safishamsi/graphify).** 코드 그래프 인덱싱, 멀티 에이전트 빌드 파이프라인, 그리고 docs/PDF/이미지/비디오에 걸친 더 넓은 지식 그래프. agentmemory는 작업을 기억하고, 이 세 프로젝트는 나머지 컨텍스트 레이어를 밝혀줍니다. 레시피와 질문 라우팅 표: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
내장 메모리 (CLAUDE.md)
-
-
-
유형
-
메모리 엔진 + MCP 서버
-
메모리 레이어 API
-
완전한 에이전트 런타임
-
개인 AI
-
메모리 API + 앱
-
팀 메모리 허브 (LLM 프록시)
-
벡터 메모리 (OSS)
-
메모리 엔진 (Oracle DB)
-
메모리 시스템
-
정적 파일
-
-
-
검색 R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
해당 없음
-
자체 보고
-
PersonaMem 76% (자체 보고)
-
~96.6% (자체 보고)
-
94.4% (자체 보고)
-
해당 없음
-
해당 없음 (grep)
-
-
-
자동 캡처
-
hook 12개 (수동 작업 없음)
-
수동 add() 호출
-
에이전트 자체 편집
-
수동
-
API 측 추출
-
프록시 가로채기 (base-URL 교체)
-
수동
-
API 추출
-
수동
-
수동 편집
-
-
-
검색
-
BM25 + Vector + Graph (RRF 융합)
-
Vector + Graph
-
Vector (archival)
-
시맨틱
-
Vector + RAG
-
4가지 자산 유형 (Chat / Skill / Wiki / CodeGraph)
-
Vector 전용
-
Vector + 시맨틱
-
감쇠 가중
-
모든 것을 컨텍스트에 로드
-
-
-
멀티 에이전트
-
MCP + REST + lease + signal
-
API (조정 없음)
-
Letta 런타임 내에서만
-
없음
-
없음
-
팀 역할 + 공유 자산
-
없음
-
스코프만 지원
-
멀티 에이전트 공유
-
에이전트별 파일
-
-
-
프레임워크 종속성
-
없음 (모든 MCP 클라이언트)
-
없음
-
높음 (Letta 사용 필수)
-
독립형
-
없음
-
프록시가 모든 모델 호출을 프론팅
-
없음
-
Oracle Database
-
없음
-
에이전트별 포맷
-
-
-
외부 의존성
-
별도 데이터베이스 서비스 불필요
-
Qdrant / pgvector
-
Postgres + 벡터 DB
-
다수
-
매니지드 클라우드
-
Docker 스택 (Core + Hub + Proxy)
-
벡터 스토어
-
Oracle AI Database
-
없음
-
없음
-
-
-
메모리 라이프사이클
-
4-tier 통합 + 감쇠 + 자동 망각
-
수동적 추출
-
에이전트 관리
-
수동
-
자동 망각
-
수동 리뷰; 자동 라우팅 진행 중
-
없음
-
명시 없음
-
감쇠 + 통합
-
수동 정리
-
-
-
토큰 효율
-
세션당 ~1,900 토큰 ($10/년)
-
통합 방식에 따라 다름
-
핵심 메모리는 컨텍스트에 상주
-
다양
-
클라우드 가격 책정
-
명시 없음
-
토큰 예산 없음
-
LLM 기반 (다양)
-
다양
-
관측 240개 기준 22K+ 토큰
-
-
-
실시간 뷰어
-
있음 (포트 3113)
-
클라우드 대시보드
-
클라우드 대시보드
-
웹 UI
-
클라우드 대시보드
-
Hub 웹 UI
-
없음
-
없음
-
없음
-
없음
-
-
-
셀프 호스팅
-
예 (기본값)
-
선택 사항
-
선택 사항
-
예
-
아니오 (클라우드 전용)
-
예 (Docker)
-
예
-
예 (Oracle DB)
-
예
-
예
-
-
-
-벤치마크 참고: agentmemory의 R@5만이 우리가 직접 측정한 결과입니다(LongMemEval-S, benchmark/COMPARISON.md에서 재현 가능). mem0와 Letta 수치는 그들이 게시한 LoCoMo 수치(다른 데이터셋)이며, MemPalace, supermemory, TencentDB (PersonaMem), oracleagentmemory 수치는 우리가 독립적으로 재현하지 않은 벤더 자체 보고 주장입니다(oracleagentmemory의 실행은 Oracle AI Database에 대해 GPT-5.5를 사용했습니다). 대략적인 비교를 위해 나란히 표시했을 뿐, 동일 데이터에 대한 정면 대결이 아닙니다. Star 수는 근사치이며 시간이 지나면서 변동합니다.
-
-알아둘 만한 **최근 진입자들**, [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md)에서 심층 비교:
-
-| 시스템 | ⭐ | 접근 방식 |
-|--------|---|-------|
-| Zep / Graphiti | 30K | 시간적 지식 그래프. 게시된 시간 쿼리 결과 중 가장 강력(LongMemEval 63.8%)하지만, 그래프가 비동기로 빌드되어 최신 사실이 지연될 수 있음 |
-| Cognee | 30K | 문서-지식 그래프 수집, Python 전용, 세션 캡처보다는 구조화된 엔티티 추출을 위해 설계됨 |
+출처를 포함한 최신 제품 및 배포 방식 비교:
-AgentMemory는 세션 캡처, 검색 가능한 프로젝트 이력, 로컬 뷰어를 함께 제공하며 처음부터 키 없이 사용할 수 있습니다.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -778,7 +570,7 @@ Verify with `curl http://localhost:3111/agentmemory/health`. Open http://localho
메모리 서버 시작: `npx -y @agentmemory/agentmemory@latest`
-#### `npx skills add`를 통한 네이티브 skill (50+ 에이전트)
+#### `npx skills add`를 통한 네이티브 skill
agentmemory는 Claude-Code 스타일의 `/SKILL.md` 형식으로 17개의 skill을 제공합니다: 호출 가능한 action skill 9개(`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`)와 에이전트가 필요할 때 로드하는 참조 skill 8개(`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`)입니다. 참조 skill은 소스에서 생성된 데이터 표를 담고 있어 절대 어긋나지 않습니다. vercel-labs의 [`skills`](https://npmjs.com/package/skills) CLI는 호출한 에이전트의 네이티브 skill 디렉터리에, 50+개의 에이전트(Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf 등)에 걸쳐 자동으로 설치합니다:
@@ -835,7 +627,7 @@ agentmemory 항목은 `mcpServers` 형태를 사용하는 모든 호스트(Curso
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks`는 공유 커스터마이징 디렉터리에 MCP와 캡처 hooks를 설치합니다. [Antigravity 설정 및 제한사항](../docs/plugins/antigravity.md) 참고. |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks`는 현재 IDE 버전과 동일한 MCP 및 hook 설정을 사용합니다. 기존 설치는 `--force`로 갱신해야 합니다. [업그레이드 노트](../docs/plugins/antigravity.md) 참고. |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro`는 사용자 수준 설정을 작성합니다. workspace 재정의는 코드 옆의 `.kiro/settings/mcp.json`에 둡니다. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp`는 표준 `mcpServers` 블록을 작성합니다. Warp는 `.claude/skills/`에서 skill도 자동으로 탐색합니다. Claude Code 플러그인을 설치하면 8개의 agentmemory skill(`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`)이 Warp의 슬래시 명령 팔레트에 네이티브로 나타납니다. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp`는 표준 `mcpServers` 블록을 작성합니다. Warp는 `.claude/skills/`에서 skill도 자동으로 탐색합니다. Claude Code 플러그인을 설치하면 8개의 agentmemory skill(`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`)이 Warp의 슬래시 명령 팔레트에 네이티브로 나타납니다. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline`은 표준 `mcpServers` 블록을 작성합니다. VS Code 확장 사용자는 Cline Settings → MCP Servers → Edit JSON을 통해 동일한 블록을 붙여넣으십시오. |
| **Continue.dev** | `~/.continue/config.yaml` (권장) 또는 `config.json` (레거시) | `agentmemory connect continue`는 설정이 없으면 `config.yaml`을 만들고 기존 `config.json`은 수정합니다. 기존 YAML은 주석과 앵커를 보존하기 위해 변경하지 않고 `mcpServers:` 아래에 붙일 블록을 출력합니다. Continue의 `mcpServers`는 배열입니다. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed`는 `context_servers`(Zed의 키이며 `mcpServers`가 아닙니다) 아래에 작성합니다. 원격 MCP 서버는 대신 `{"url": "..."}`로 연결할 수 있습니다. |
@@ -932,7 +724,6 @@ Session 2: "Now add rate limiting"
|---|---|---|
| 관리 | 수동으로 관리하는 지침 | 색인된 세션 이력과 메모리 |
| 검색 | 모든 것을 컨텍스트에 로드 | BM25 + vector + graph (top-K만) |
-| 토큰 비용 | 관측 240개 기준 22K+ | ~1,900 토큰 (92% 감소) |
| 에이전트 간 | 에이전트별 파일 | MCP + REST (모든 에이전트) |
| 조정 | 없음 | lease, signal, action, routine |
| 관측 가능성 | 파일을 수동으로 읽기 | :3113의 실시간 뷰어 |
@@ -945,7 +736,7 @@ Session 2: "Now add rate limiting"
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1030,7 +821,7 @@ Reciprocal Rank Fusion(RRF, k=60)으로 융합되고, 세션별로 다변화됩
벡터 인덱스가 채워져 있으면, (`memory_recall` 뒤에 있는) `mem::search`는 하이브리드 BM25 + vector 랭커를 사용합니다. 임베딩이 없으면 BM25를 사용합니다. `smart-search`는 그래프 데이터가 존재하면 키 없는 모드에서도 구조적 그래프 매치를 추가로 융합할 수 있습니다. Lesson 리콜은 쿼리마다 전체 코퍼스를 스캔하는 대신 전용 인메모리 BM25 인덱스에서 동작합니다. 대체된 메모리 버전은 모든 리콜 경로에서 제외되며, 버전 체인이 그 기록을 보존합니다.
-벡터는 크래시나 강제 종료에도 유지됩니다. 벡터 인덱스는 최소 `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS`(10분) 간격으로 버킷 단위로 저장됩니다. 그 사이에 추가되거나 제거된 모든 벡터는 상태 저장소의 작은 pending 로그에도 즉시 기록되며, 다음 시작 시 임베딩 프로바이더를 호출하지 않고 이를 재생합니다. 저장이 성공할 때마다 로그가 비워집니다. 재생 후에도 벡터가 없는 문서는 `AGENTMEMORY_VECTOR_BACKFILL_MAX`(500)개 단위 배치로 백그라운드에서 다시 임베딩되며, 더 이상 남지 않을 때까지 계속됩니다. 중단된 백필은 다음 시작 시 이어서 진행됩니다. `/agentmemory/status`와 뷰어는 pending 로그 크기와 백필 상태를 보여줍니다. 키 없는 설치는 아무것도 기록하지 않습니다.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25는 기본적으로 그리스어, 키릴 문자, 히브리어, 아랍어, 강세 부호가 있는 라틴 문자를 토크나이즈합니다. 중국어 / 일본어 / 한국어 메모리의 경우, 선택적 세그멘터(`npm install @node-rs/jieba tiny-segmenter`)를 설치해 CJK 런을 단어 수준 토큰으로 분할하십시오. 설치하지 않으면 agentmemory는 전체 런 토크나이제이션으로 soft fallback하고 stderr에 일회성 힌트를 출력합니다.
@@ -1046,9 +837,9 @@ EMBEDDING_PROVIDER=local
| 프로바이더 | 모델 | 비용 | 참고 |
|---|---|---|---|
-| **Local (권장 opt-in)** | `all-MiniLM-L6-v2` | 무료 | 첫 모델 다운로드 이후 온디바이스 실행, BM25 단독 대비 recall +8pp |
+| **Local (권장 opt-in)** | `all-MiniLM-L6-v2` | 무료 | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | 무료 티어 | 100+개 언어, 768/1536/3072차원(MRL), 2048 토큰 입력. `text-embedding-004`를 대체함([지원 종료, 2026년 1월 14일 중단](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | 최고 품질 |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | 유료 | 코드에 최적화됨 |
| Cohere | `embed-english-v3.0` | 무료 체험 | 범용 |
| OpenRouter | 모든 모델 | 다양함 | 멀티 모델 프록시 |
diff --git a/READMEs/README.nb-NO.md b/READMEs/README.nb-NO.md
index b8bc28c547..620048b461 100644
--- a/READMEs/README.nb-NO.md
+++ b/READMEs/README.nb-NO.md
@@ -66,8 +66,6 @@
-
-> Embedding-modell: `all-MiniLM-L6-v2` (lokal, gratis, ingen API-nøkkel). Fullstendige rapporter: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Konkurrentsammenligning: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) dekker agentmemory mot mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Reprodusér lokalt:** [`eval/README.md`](../eval/README.md), en adapter-plugbar harness for LongMemEval `_s` (offentlig 500 spørsmål) + `coding-agent-life-v1` (eget korpus på 15 sesjoner). Grep-, vektor- og agentmemory-adaptere scores side ved side, NDJSON-utdata, publiserte scorekort havner i [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Fungerer godt sammen med [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything), og [Graphify](https://github.com/safishamsi/graphify).** Kodegrafindeksering, flerstegs byggepipeliner, og bredere kunnskapsgrafer over dokumenter/PDF-er/bilder/videoer. agentmemory husker arbeidet; disse tre prosjektene lyser opp resten av kontekstlaget. Oppskrifter + spørsmål-rutingstabell: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-Merk om ytelsestester: kun agentmemorys R@5 er vårt eget målte resultat (LongMemEval-S, reproduserbart fra benchmark/COMPARISON.md). Tallene for mem0 og Letta er deres publiserte LoCoMo-tall (et annet datasett); tallene for MemPalace, supermemory, TencentDB (PersonaMem) og oracleagentmemory er leverandørers selvrapporterte påstander vi ikke har reprodusert selv (oracleagentmemorys kjøring brukte GPT-5.5 mot en Oracle AI Database). Vist side ved side kun som et grovt sammenligningsgrunnlag, ikke en direkte sammenligning på identiske data. Stjernetall er tilnærmede og endrer seg over tid.
-
-**Nyere aktører** verdt å kjenne til, sammenlignet i detalj i [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| System | ⭐ | Vinkling |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporal kunnskapsgraf; de sterkeste publiserte resultatene på temporale søk (LongMemEval 63.8%), men grafen bygges asynkront så ferske fakta kan ligge etter |
-| Cognee | 30K | Dokument-til-kunnskapsgraf-innlasting, kun Python, bygget for strukturert entitetsuttrekk snarere enn sesjonsfangst |
+Oppdatert sammenligning av produkter og driftsalternativer med kilder:
-AgentMemory kombinerer registrering fra kodeagenter, søkbar prosjekthistorikk og en lokal viewer. Du kan arbeide uten nøkkel fra første start.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -777,7 +569,7 @@ Full veiledning: [`integrations/hermes/`](../integrations/hermes/)
Start minneserveren: `npx -y @agentmemory/agentmemory@latest`
-#### Native skills via `npx skills add` (50+ agenter)
+#### Native skills via `npx skills add`
agentmemory leveres med 17 skills i Claude-Code-stilens `/SKILL.md`-format: 9 skills som kan kalles (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) og 8 referanseskills agenten laster ved behov (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Referanseskillene har datatabeller generert fra kildekoden, så de aldri går ut av synk. CLI-en [`skills`](https://npmjs.com/package/skills) fra vercel-labs installerer dem automatisk inn i den kallende agentens native skill-katalog over 50+ agenter (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, med flere):
@@ -834,7 +626,7 @@ agentmemory-oppføringen er **samme MCP-server-blokk** på tvers av alle verter
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` installerer MCP og fangst-hooks i den delte tilpasningskatalogen. Se [Antigravity-oppsett og begrensninger](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` bruker samme MCP- og hook-konfigurasjon som nåværende IDE-versjoner. Eksisterende installasjoner bør oppdateres med `--force`; se [oppgraderingsnotatene](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` skriver brukernivå-konfigurasjonen. Arbeidsområde-overstyringer går i `.kiro/settings/mcp.json` ved siden av koden din. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` skriver standard `mcpServers`-blokk. Warp oppdager også automatisk skills fra `.claude/skills/`; når Claude Code-pluginen er installert vises de 8 agentmemory-skillene (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) naturlig i Warps slash-kommando-palett. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` skriver standard `mcpServers`-blokk. Warp oppdager også automatisk skills fra `.claude/skills/`; når Claude Code-pluginen er installert vises de 9 agentmemory-skillene (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) naturlig i Warps slash-kommando-palett. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` skriver standard `mcpServers`-blokk. VS Code-utvidelsesbrukere: lim inn samme blokk via Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (foretrukket) eller `config.json` (eldre) | `agentmemory connect continue` oppretter en manglende `config.yaml` eller endrer eksisterende `config.json`. Hvis YAML-filen finnes, vises blokken for `mcpServers:` slik at kommentarer og ankere bevares. Continue bruker et array, ikke et objekt. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` skriver under `context_servers` (Zeds nøkkel, IKKE `mcpServers`). Eksterne MCP-servere kan kobles til via `{"url": "..."}` i stedet. |
@@ -931,7 +723,6 @@ Hver AI-kodeagent leveres med innebygd hukommelse: Claude Code har `MEMORY.md`,
|---|---|---|
| Vedlikehold | Manuelt vedlikeholdte instruksjoner | Indeksert økthistorikk og minner |
| Søk | Laster alt inn i konteksten | BM25 + vektor + graf (kun topp-K) |
-| Tokenkostnad | 22K+ ved 240 observasjoner | ~1 900 tokens (92 % mindre) |
| Tvers av agenter | Filer per agent | MCP + REST (hvilken som helst agent) |
| Koordinering | Ingen | Leaser, signaler, handlinger, rutiner |
| Observerbarhet | Les filer manuelt | Viewer i realtid på :3113 |
@@ -944,7 +735,7 @@ Hver AI-kodeagent leveres med innebygd hukommelse: Claude Code har `MEMORY.md`,
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1029,7 +820,7 @@ Fusjonert med Reciprocal Rank Fusion (RRF, k=60) og sesjonsdiversifisert (maks 3
Når en vektorindeks er fylt, bruker `mem::search` (bak `memory_recall`) den hybride BM25 + vektor-rangereren. Uten embeddings bruker den BM25. `smart-search` kan i tillegg slå sammen strukturelle grafmatcher når grafdata finnes, også i nøkkelfri modus. Gjenkalling av læringspunkter kjører på en dedikert BM25-indeks i minnet i stedet for å skanne hele korpuset per søk. Erstattede minneversjoner er utelatt fra hver gjenkallingsvei; versjonskjeden beholder historikken deres.
-Vektorer overlever en krasj eller tvungen avslutning. Vektorindeksen lagres i bøtter minst hvert `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minutter). Hver vektor som legges til eller fjernes i mellomtiden skrives også umiddelbart til en liten ventende logg i state-lageret, og neste oppstart spiller den av igjen uten å kalle embedding-leverandøren. Hver vellykkede lagring tømmer loggen. Dokumenter som fortsatt ikke har en vektor etter avspillingen, re-embeddes i bakgrunnen i batcher av `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) til ingen er igjen, og en etterfylling som stoppes fortsetter ved neste oppstart. `/agentmemory/status` og viewer viser størrelsen på den ventende loggen og etterfyllingsstatusen. Nøkkelfrie installasjoner skriver ingenting.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokeniserer gresk, kyrillisk, hebraisk, arabisk, og aksentuert latin ut av boksen. For kinesiske/japanske/koreanske minner, installer de valgfrie segmenterne (`npm install @node-rs/jieba tiny-segmenter`) for å dele opp CJK-sekvenser i ord-nivå-tokens; uten dem faller agentmemory mykt tilbake til hele-sekvens-tokenisering og skriver et engangs-hint til stderr.
@@ -1045,9 +836,9 @@ Den vanlige npm-installasjonen inkluderer den valgfrie `@huggingface/transformer
| Leverandør | Modell | Kostnad | Merknader |
|---|---|---|---|
-| **Lokal (anbefalt opt-in)** | `all-MiniLM-L6-v2` | Gratis | På enheten etter den første modellnedlastingen, +8 prosentpoeng gjenkalling over kun-BM25 |
+| **Lokal (anbefalt opt-in)** | `all-MiniLM-L6-v2` | Gratis | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Gratis nivå | 100+ språk, 768/1536/3072 dimensjoner (MRL), 2048-tokens inndata. Erstatter `text-embedding-004` ([avviklet, stengt 14. januar 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Høyest kvalitet |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Betalt | Optimalisert for kode |
| Cohere | `embed-english-v3.0` | Gratis prøveperiode | Generell bruk |
| OpenRouter | Hvilken som helst modell | Varierer | Flermodell-proxy |
diff --git a/READMEs/README.nl-NL.md b/READMEs/README.nl-NL.md
index 864d483c54..835d336cdd 100644
--- a/READMEs/README.nl-NL.md
+++ b/READMEs/README.nl-NL.md
@@ -66,8 +66,6 @@
-
-
@@ -329,223 +327,17 @@ Laatste release notes: [CHANGELOG.md](../CHANGELOG.md).
-
-
-
-
-### Retrieval-nauwkeurigheid
-
-**coding-agent-life-v1** (eigen corpus, reproduceerbaar in een sandbox)
-
-| Adapter | P@5 | R@5 | Top-5-hitrate | p50-latentie |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep-baseline | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-100% top-5-hitrate op het **P@5-wiskundige plafond** voor dit corpus (0.240, zie scorecard). Hybride haalt elke gold session op; grep mist 1 van 2 gold-resultaten bij de multi-session temporele query. De winst zit in **recall + temporeel**, niet in geaggregeerde precisie. Deze benchmark is klein en gold-sparse; de grotere LongMemEval-S hieronder onderscheidt beter. Volledige uitsplitsing per type + correctienotitie: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 vragen)
+Actuele zoekresultaten, definities van meetwaarden en stappen om de test te herhalen:
-| Systeem | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| BM25-only fallback | 86.2% | 94.6% | 71.5% |
-
-
-
-> Embeddingmodel: `all-MiniLM-L6-v2` (lokaal, gratis, geen API-key nodig). Volledige rapporten: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Vergelijking met concurrenten: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md), met agentmemory tegenover mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee en Hippo.
-
-**Lokaal reproduceren:** [`eval/README.md`](../eval/README.md), een harness met inpluggbare adapters voor LongMemEval `_s` (publieke set van 500 vragen) + `coding-agent-life-v1` (eigen corpus van 15 sessies). Grep-, vector- en agentmemory-adapters scoren naast elkaar, NDJSON-output, gepubliceerde scorecards komen terecht in [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Combineert goed met [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) en [Graphify](https://github.com/safishamsi/graphify).** Code-graph-indexering, multi-agent build-pipelines en bredere knowledge graphs over docs / PDF's / afbeeldingen / video's. agentmemory onthoudt het werk; die drie projecten verlichten de rest van de context-laag. Recepten + tabel voor vraagroutering: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Ingebouwd (CLAUDE.md)
-
-
-
Type
-
Memory engine + MCP-server
-
Memory-laag-API
-
Volledige agent-runtime
-
Persoonlijke AI
-
Memory-API + app
-
Team-memory-hub (LLM-proxy)
-
Vector-memory (OSS)
-
Memory engine (Oracle DB)
-
Memory-systeem
-
Statisch bestand
-
-
-
Retrieval R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Zelf gerapporteerd
-
PersonaMem 76% (zelf gerapporteerd)
-
~96.6% (zelf gerapporteerd)
-
94.4% (zelf gerapporteerd)
-
N/A
-
N/A (grep)
-
-
-
Auto-capture
-
12 hooks (geen handmatige moeite)
-
Handmatige add()-aanroepen
-
Agent bewerkt zelf
-
Handmatig
-
Extractie aan de API-kant
-
Proxy-interceptie (base-URL wisselen)
-
Handmatig
-
API-extractie
-
Handmatig
-
Handmatig bewerken
-
-
-
Zoeken
-
BM25 + Vector + Graph (RRF-fusie)
-
Vector + Graph
-
Vector (archief)
-
Semantisch
-
Vector + RAG
-
4 asset-types (Chat / Skill / Wiki / CodeGraph)
-
Alleen vector
-
Vector + semantisch
-
Decay-gewogen
-
Laadt alles in de context
-
-
-
Multi-agent
-
MCP + REST + leases + signals
-
API (geen coördinatie)
-
Alleen binnen Letta-runtime
-
Nee
-
Nee
-
Teamrollen + gedeelde assets
-
Nee
-
Alleen scoped
-
Gedeeld tussen meerdere agents
-
Bestanden per agent
-
-
-
Framework lock-in
-
Geen (elke MCP-client)
-
Geen
-
Hoog (moet Letta gebruiken)
-
Standalone
-
Geen
-
Proxy staat voor elke model-aanroep
-
Geen
-
Oracle Database
-
Geen
-
Formaat per agent
-
-
-
Externe dependencies
-
Geen afzonderlijke databaseservice nodig
-
Qdrant / pgvector
-
Postgres + vector-db
-
Meerdere
-
Managed cloud
-
Docker-stack (Core + Hub + Proxy)
-
Vector store
-
Oracle AI Database
-
Geen
-
Geen
-
-
-
Memory lifecycle
-
4-laags consolidatie + decay + auto-forget
-
Passieve extractie
-
Beheerd door agent
-
Handmatig
-
Auto-forget
-
Handmatige review; auto-routing in ontwikkeling
-
Geen
-
Niet vermeld
-
Decay + consolidatie
-
Handmatig opschonen
-
-
-
Tokenefficiëntie
-
~1,900 tokens/sessie ($10/jaar)
-
Varieert per integratie
-
Core memory in context
-
Varieert
-
Cloud pricing
-
Niet vermeld
-
Geen tokenbudget
-
LLM-ondersteund (varieert)
-
Varieert
-
22K+ tokens bij 240 observaties
-
-
-
Realtime viewer
-
Ja (poort 3113)
-
Cloud dashboard
-
Cloud dashboard
-
Web UI
-
Cloud dashboard
-
Hub web UI
-
Nee
-
Nee
-
Nee
-
Nee
-
-
-
Self-hosted
-
Ja (standaard)
-
Optioneel
-
Optioneel
-
Ja
-
Nee (alleen cloud)
-
Ja (Docker)
-
Ja
-
Ja (Oracle DB)
-
Ja
-
Ja
-
-
-
-Opmerking over de benchmark: alleen de R@5 van agentmemory is ons eigen gemeten resultaat (LongMemEval-S, reproduceerbaar via benchmark/COMPARISON.md). De cijfers van mem0 en Letta zijn hun gepubliceerde LoCoMo-cijfers (een andere dataset); de cijfers van MemPalace, supermemory, TencentDB (PersonaMem) en oracleagentmemory zijn door de leverancier zelf gerapporteerde claims die we niet onafhankelijk hebben gereproduceerd (de run van oracleagentmemory gebruikte GPT-5.5 tegen een Oracle AI Database). Naast elkaar getoond voor een ruwe inschatting, niet als head-to-head op identieke data. Stertallen zijn benaderend en veranderen in de tijd.
-
-**Nieuwere spelers** die het kennen waard zijn, in detail vergeleken in [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Systeem | ⭐ | Invalshoek |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporele knowledge graph; sterkste gepubliceerde resultaten voor temporele queries (LongMemEval 63.8%), maar de graph wordt asynchroon opgebouwd, waardoor verse feiten kunnen achterblijven |
-| Cognee | 30K | Document-naar-knowledge-graph-ingestie, alleen Python, gebouwd voor gestructureerde entiteitsextractie in plaats van sessie-capture |
+Bijgewerkte vergelijking van producten en implementatieopties, met bronnen:
-AgentMemory combineert vastlegging door programmeeragents, doorzoekbare projectgeschiedenis en een lokale viewer. Je kunt vanaf de eerste start zonder sleutel werken.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -775,7 +567,7 @@ Volledige handleiding: [`integrations/hermes/`](../integrations/hermes/)
Start de memory server: `npx -y @agentmemory/agentmemory@latest`
-#### Native skills via `npx skills add` (50+ agents)
+#### Native skills via `npx skills add`
agentmemory levert 17 skills in het Claude-Code-achtige `/SKILL.md`-formaat: 9 aanroepbare actie-skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) en 8 referentieskills die de agent op aanvraag laadt (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). De referentieskills bevatten datatabellen die uit de bron gegenereerd zijn, dus ze lopen nooit uit de pas. De [`skills`](https://npmjs.com/package/skills)-CLI van vercel-labs installeert ze automatisch in de native skills-map van de aanroepende agent, verspreid over 50+ agents (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, en meer):
@@ -832,7 +624,7 @@ De agentmemory-entry is hetzelfde **MCP-serverblok** op elke host die de `mcpSer
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` installeert MCP en capture-hooks in de gedeelde customization-map. Zie [Antigravity-setup en -beperkingen](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` gebruikt dezelfde MCP- en hookconfiguratie als de huidige IDE-versies. Bestaande installaties moeten verversen met `--force`; zie de [upgrade-notities](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` schrijft de config op gebruikersniveau. Workspace-overrides komen in `.kiro/settings/mcp.json` naast je code. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` schrijft het standaard `mcpServers`-blok. Warp ontdekt ook automatisch skills vanuit `.claude/skills/`; zodra de Claude Code-plugin is geïnstalleerd, verschijnen de 8 agentmemory-skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) native in het slashcommando-palet van Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` schrijft het standaard `mcpServers`-blok. Warp ontdekt ook automatisch skills vanuit `.claude/skills/`; zodra de Claude Code-plugin is geïnstalleerd, verschijnen de 9 agentmemory-skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) native in het slashcommando-palet van Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` schrijft het standaard `mcpServers`-blok. Gebruikers van de VS Code-extensie: plak hetzelfde blok via Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (voorkeur) of `config.json` (legacy) | `agentmemory connect continue` maakt een ontbrekende `config.yaml` of wijzigt een bestaande `config.json`. Bij bestaande YAML toont de adapter het blok voor `mcpServers:` zodat comments en anchors behouden blijven. Continue gebruikt een array, geen object. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` schrijft onder `context_servers` (de key van Zed, NIET `mcpServers`). Remote MCP-servers kunnen in plaats daarvan gekoppeld worden via `{"url": "..."}`. |
@@ -929,7 +721,6 @@ Elke AI coding agent levert ingebouwd geheugen: Claude Code heeft `MEMORY.md`, C
|---|---|---|
| Onderhoud | Handmatig bijgehouden instructies | Geïndexeerde sessiegeschiedenis en herinneringen |
| Zoeken | Laadt alles in de context | BM25 + vector + graph (alleen top-K) |
-| Tokenkosten | 22K+ bij 240 observaties | ~1,900 tokens (92% minder) |
| Cross-agent | Bestanden per agent | MCP + REST (elke agent) |
| Coördinatie | Geen | Leases, signals, actions, routines |
| Observability | Bestanden handmatig lezen | Realtime viewer op :3113 |
@@ -942,7 +733,7 @@ Elke AI coding agent levert ingebouwd geheugen: Claude Code heeft `MEMORY.md`, C
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1027,7 +818,7 @@ Samengevoegd met Reciprocal Rank Fusion (RRF, k=60) en gediversifieerd per sessi
Wanneer een vector-index gevuld is, gebruikt `mem::search` (achter `memory_recall`) de hybride BM25 + vector-ranker. Zonder embeddings gebruikt het BM25. `smart-search` kan daarnaast structurele graph-matches samenvoegen wanneer er graph-data bestaat, ook in keyless-modus. Lesson-recall draait op een eigen in-memory BM25-index in plaats van het hele corpus per query te doorzoeken. Vervangen memory-versies worden uitgesloten van elk recall-pad; de versieketen behoudt hun geschiedenis.
-Vectoren overleven een crash of force-kill. De vector-index wordt in batches opgeslagen, maximaal elke `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minuten). Elke vector die er tussendoor bijkomt of verdwijnt, wordt ook meteen geschreven naar een klein pending-log in de state store, en de volgende start speelt dit af zonder de embedding-provider aan te roepen. Elke succesvolle save maakt het log leeg. Documenten die na het afspelen nog geen vector hebben, worden op de achtergrond opnieuw ge-embed in batches van `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) totdat er geen meer over zijn, en een backfill die gestopt wordt, gaat bij de volgende start verder. `/agentmemory/status` en de viewer tonen de grootte van het pending-log en de status van de backfill. Keyless-installaties schrijven niets.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokeniseert standaard Grieks, Cyrillisch, Hebreeuws, Arabisch en Latijn met accenten. Voor Chinese / Japanse / Koreaanse memories installeer je de optionele segmenters (`npm install @node-rs/jieba tiny-segmenter`) om CJK-reeksen op te splitsen in tokens op woordniveau; zonder deze valt agentmemory soft terug op tokenisatie van de hele reeks en print het eenmalig een hint naar stderr.
@@ -1043,9 +834,9 @@ De normale npm-installatie bevat de optionele `@huggingface/transformers`-runtim
| Provider | Model | Kosten | Opmerkingen |
|---|---|---|---|
-| **Lokaal (aanbevolen opt-in)** | `all-MiniLM-L6-v2` | Gratis | Op het eigen apparaat na de eerste modeldownload, +8pp recall t.o.v. alleen BM25 |
+| **Lokaal (aanbevolen opt-in)** | `all-MiniLM-L6-v2` | Gratis | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Gratis tier | 100+ talen, 768/1536/3072 dims (MRL), input tot 2048 tokens. Vervangt `text-embedding-004` ([deprecated, uitgefaseerd op 14 jan. 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Hoogste kwaliteit |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Betaald | Geoptimaliseerd voor code |
| Cohere | `embed-english-v3.0` | Gratis trial | Algemeen inzetbaar |
| OpenRouter | Elk model | Varieert | Multi-model-proxy |
diff --git a/READMEs/README.pl-PL.md b/READMEs/README.pl-PL.md
index 5bc508604d..b19867232a 100644
--- a/READMEs/README.pl-PL.md
+++ b/READMEs/README.pl-PL.md
@@ -66,8 +66,6 @@
-
-
@@ -331,223 +329,17 @@ Najnowsze informacje o wydaniach: [CHANGELOG.md](../CHANGELOG.md).
-
-
-
-
-### Dokładność wyszukiwania
-
-**coding-agent-life-v1** (korpus własny, odtwarzalny w sandboksie)
-
-| Adapter | P@5 | R@5 | Trafienia w top-5 | opóźnienie p50 |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep baseline | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-100% trafień w top-5 przy **matematycznym suficie P@5** dla tego korpusu (0.240, zobacz scorecard). Hybrid odzyskuje każdą złotą sesję; grep pomija 1 z 2 złotych sesji w zapytaniu temporalnym obejmującym wiele sesji. Przewaga dotyczy **recall + czasu**, a nie zagregowanej precyzji. Ten benchmark jest mały i ma mało danych złotych; większy LongMemEval-S poniżej różnicuje lepiej. Pełny podział per typ + uwaga o korekcie: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 pytań)
+Aktualne wyniki wyszukiwania, definicje metryk i kroki odtworzenia testu:
-| System | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| Fallback tylko BM25 | 86.2% | 94.6% | 71.5% |
-
-
-
-Uwaga dotycząca benchmarku: tylko wynik R@5 agentmemory jest naszym własnym zmierzonym wynikiem (LongMemEval-S, odtwarzalny na podstawie benchmark/COMPARISON.md). Wartości mem0 i Letta to ich opublikowane liczby LoCoMo (inny zbiór danych); wartości MemPalace, supermemory, TencentDB (PersonaMem) i oracleagentmemory to deklaracje producentów, których nie odtworzyliśmy niezależnie (uruchomienie oracleagentmemory używało GPT-5.5 względem Oracle AI Database). Zestawione obok siebie tylko orientacyjnie, nie jako test głowa w głowę na identycznych danych. Liczby gwiazdek są przybliżone i zmieniają się w czasie.
-
-**Nowsi uczestnicy** warci uwagi, porównani szerzej w [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| System | ⭐ | Podejście |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporalny graf wiedzy; najsilniejsze opublikowane wyniki zapytań temporalnych (LongMemEval 63.8%), ale graf buduje się asynchronicznie, więc świeże fakty mogą się spóźniać |
-| Cognee | 30K | Ingestia dokumentów do grafu wiedzy, tylko Python, zbudowane do strukturalnej ekstrakcji encji, a nie przechwytywania sesji |
+Aktualne porównanie produktów i opcji wdrożenia wraz ze źródłami:
-AgentMemory łączy przechwytywanie sesji, przeszukiwalną historię projektu i lokalny podgląd, z trybem bez klucza od pierwszego uruchomienia.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -784,7 +576,7 @@ Pełny przewodnik: [`integrations/hermes/`](../integrations/hermes/)
Uruchom serwer pamięci: `npx -y @agentmemory/agentmemory@latest`
-#### Natywne skille przez `npx skills add` (50+ agentów)
+#### Natywne skille przez `npx skills add`
agentmemory dostarcza 17 skilli w formacie `/SKILL.md` w stylu Claude Code: 9 wywoływalnych skilli akcji (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) oraz 8 skilli referencyjnych wczytywanych przez agenta na żądanie (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Skille referencyjne zawierają tabele danych generowane ze źródła, więc nigdy się nie rozjeżdżają. CLI [`skills`](https://npmjs.com/package/skills) od vercel-labs automatycznie instaluje je w natywnym katalogu skilli wywołującego agenta, obejmując ponad 50 agentów (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf i inne):
@@ -842,7 +634,7 @@ Wpis agentmemory to **ten sam blok serwera MCP** we wszystkich hostach, które u
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` instaluje MCP i hooki przechwytywania we współdzielonym katalogu personalizacji. Zobacz [Konfiguracja i ograniczenia Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` używa tej samej konfiguracji MCP i hooków, jak aktualne wersje IDE. Istniejące instalacje powinny odświeżyć się za pomocą `--force`; zobacz [uwagi dotyczące aktualizacji](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` zapisuje konfigurację na poziomie użytkownika. Nadpisania na poziomie workspace idą do `.kiro/settings/mcp.json` obok twojego kodu. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` zapisuje standardowy blok `mcpServers`. Warp automatycznie wykrywa też skille z `.claude/skills/`; gdy zainstalowany jest plugin Claude Code, 8 skilli agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) pojawia się natywnie w palecie komend ukośnikowych Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` zapisuje standardowy blok `mcpServers`. Warp automatycznie wykrywa też skille z `.claude/skills/`; gdy zainstalowany jest plugin Claude Code, 9 skilli agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) pojawia się natywnie w palecie komend ukośnikowych Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` zapisuje standardowy blok `mcpServers`. Użytkownicy rozszerzenia VS Code: wklej ten sam blok przez Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (preferowane) lub `config.json` (starsze) | `agentmemory connect continue` tworzy `config.yaml`, jeśli konfiguracji nie ma, lub modyfikuje istniejący `config.json`. Dla istniejącego YAML wypisuje blok do wklejenia pod `mcpServers:`, zachowując komentarze i kotwice. Continue używa tablicy, nie obiektu. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` zapisuje pod `context_servers` (klucz Zed, a NIE `mcpServers`). Zdalne serwery MCP można podłączyć zamiast tego przez `{"url": "..."}`. |
@@ -925,7 +717,6 @@ Każdy kodujący agent AI ma wbudowaną pamięć: Claude Code ma `MEMORY.md`, Cu
|---|---|---|
| Utrzymanie | Ręcznie zarządzane instrukcje | Indeksowana historia sesji i wspomnienia |
| Wyszukiwanie | Wczytuje wszystko do kontekstu | BM25 + wektor + graf (tylko top-K) |
-| Koszt tokenów | 22K+ przy 240 obserwacjach | ~1,900 tokenów (92% mniej) |
| Między agentami | Pliki per agent | MCP + REST (każdy agent) |
| Koordynacja | Brak | Leasy, sygnały, akcje, rutyny |
| Obserwowalność | Ręczne czytanie plików | Podgląd w czasie rzeczywistym na :3113 |
@@ -938,7 +729,7 @@ Każdy kodujący agent AI ma wbudowaną pamięć: Claude Code ma `MEMORY.md`, Cu
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1023,7 +814,7 @@ Wyszukiwanie w trzech strumieniach, łączące trzy sygnały:
Gdy indeks wektorowy jest wypełniony, `mem::search` (stojące za `memory_recall`) używa hybrydowego rankera BM25 + wektor. Bez embeddingów używa BM25. `smart-search` może dodatkowo łączyć strukturalne dopasowania z grafu, gdy dane grafu istnieją, również w trybie bez klucza. Przywołanie lekcji działa na dedykowanym indeksie BM25 w pamięci, zamiast skanować cały korpus przy każdym zapytaniu. Zastąpione wersje pamięci są wykluczone z każdej ścieżki przywołania; łańcuch wersji zachowuje ich historię.
-Wektory przetrwają awarię lub wymuszone zabicie procesu. Indeks wektorowy jest zapisywany w paczkach (buckets) nie częściej niż co `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minut). Każdy wektor dodany lub usunięty w międzyczasie jest też natychmiast zapisywany do małego dziennika oczekujących w magazynie stanu, a przy kolejnym starcie jest odtwarzany bez wywoływania dostawcy embeddingów. Każdy udany zapis opróżnia ten dziennik. Dokumenty, które po odtworzeniu wciąż nie mają wektora, są ponownie embeddowane w tle w paczkach po `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500), aż żaden nie zostanie, a zatrzymane uzupełnianie kontynuuje się przy następnym starcie. `/agentmemory/status` i podgląd pokazują rozmiar dziennika oczekujących i stan uzupełniania. Instalacje bez klucza nic nie zapisują.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 od razu tokenizuje grekę, cyrylicę, hebrajski, arabski i akcentowaną łacinę. Dla wspomnień w chińskim / japońskim / koreańskim zainstaluj opcjonalne segmentery (`npm install @node-rs/jieba tiny-segmenter`), aby dzielić ciągi CJK na tokeny na poziomie słów; bez nich agentmemory łagodnie przechodzi na tokenizację całych ciągów i jednorazowo wypisuje wskazówkę na stderr.
@@ -1039,9 +830,9 @@ Zwykła instalacja npm zawiera opcjonalne środowisko uruchomieniowe `@huggingfa
| Dostawca | Model | Koszt | Uwagi |
|---|---|---|---|
-| **Lokalny (zalecany opt-in)** | `all-MiniLM-L6-v2` | Bezpłatny | Lokalnie po pierwszym pobraniu modelu, +8pp recall względem samego BM25 |
+| **Lokalny (zalecany opt-in)** | `all-MiniLM-L6-v2` | Bezpłatny | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Darmowy poziom | 100+ języków, wymiary 768/1536/3072 (MRL), wejście do 2048 tokenów. Zastępuje `text-embedding-004` ([przestarzały, wyłączenie 14 stycznia 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Najwyższa jakość |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Płatny | Zoptymalizowany pod kod |
| Cohere | `embed-english-v3.0` | Darmowy okres próbny | Ogólnego przeznaczenia |
| OpenRouter | Każdy model | Zależnie | Proxy dla wielu modeli |
diff --git a/READMEs/README.pt-BR.md b/READMEs/README.pt-BR.md
index 864862b69d..8630533003 100644
--- a/READMEs/README.pt-BR.md
+++ b/READMEs/README.pt-BR.md
@@ -66,8 +66,6 @@
-
-### Precisão de recuperação
-
-**coding-agent-life-v1** (corpus interno, reproduzível em sandbox)
-
-| Adaptador | P@5 | R@5 | Taxa de acerto top-5 | Latência p50 |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep baseline | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-Taxa de acerto top-5 de 100% no **teto matemático de P@5** deste corpus (0.240, veja a scorecard). O híbrido recupera todas as sessões gold; o grep perde 1 de 2 golds na query temporal multi-sessão. O ganho é **recall + temporal**, não precisão agregada. Este benchmark é pequeno e esparso em golds; o LongMemEval-S maior abaixo diferencia melhor. Detalhamento completo por tipo + nota de correção: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 perguntas)
+Resultados atuais de recuperação, definições das métricas e etapas de reprodução:
-| Sistema | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| BM25-only fallback | 86.2% | 94.6% | 71.5% |
-
-
-
-> Modelo de embedding: `all-MiniLM-L6-v2` (local, gratuito, sem API key). Relatórios completos: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Comparativo com concorrentes: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) cobrindo agentmemory vs mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Reproduza localmente:** [`eval/README.md`](../eval/README.md), um harness com adaptadores plugáveis para LongMemEval `_s` (500-Q públicas) e `coding-agent-life-v1` (corpus interno de 15 sessões). Adaptadores grep / vector / agentmemory são pontuados lado a lado, saída em NDJSON, e as scorecards publicadas ficam em [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Combina com [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) e [Graphify](https://github.com/safishamsi/graphify).** Indexação de grafos de código, pipelines de build multiagente e grafos de conhecimento mais amplos sobre docs / PDFs / imagens / vídeos. agentmemory lembra do trabalho; esses três projetos iluminam o resto da camada de contexto. Recipes e tabela de roteamento por pergunta: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Built-in (CLAUDE.md)
-
-
-
Tipo
-
Engine de memória + servidor MCP
-
API de camada de memória
-
Runtime de agente completo
-
IA pessoal
-
API de memória + app
-
Hub de memória de time (proxy de LLM)
-
Memória vetorial (OSS)
-
Engine de memória (Oracle DB)
-
Sistema de memória
-
Arquivo estático
-
-
-
Retrieval R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Autorreportado
-
PersonaMem 76% (autorreportado)
-
~96.6% (autorreportado)
-
94.4% (autorreportado)
-
N/A
-
N/A (grep)
-
-
-
Captura automática
-
12 hooks (esforço manual zero)
-
Chamadas manuais a add()
-
O agente se autoedita
-
Manual
-
Extração no lado da API
-
Interceptação por proxy (troca de base-URL)
-
Manual
-
Extração via API
-
Manual
-
Edição manual
-
-
-
Busca
-
BM25 + Vector + Graph (fusão RRF)
-
Vector + Graph
-
Vector (archival)
-
Semântica
-
Vector + RAG
-
4 tipos de asset (Chat / Skill / Wiki / CodeGraph)
-
Somente vector
-
Vector + semântica
-
Ponderada por decaimento
-
Carrega tudo no contexto
-
-
-
Multiagente
-
MCP + REST + leases + signals
-
API (sem coordenação)
-
Somente dentro do runtime do Letta
-
Não
-
Não
-
Papéis de time + assets compartilhados
-
Não
-
Somente com escopo
-
Compartilhado multiagente
-
Arquivos por agente
-
-
-
Dependência de framework
-
Nenhuma (qualquer cliente MCP)
-
Nenhuma
-
Alta (precisa usar Letta)
-
Standalone
-
Nenhuma
-
Proxy na frente de toda chamada de modelo
-
Nenhuma
-
Oracle Database
-
Nenhuma
-
Formato por agente
-
-
-
Dependências externas
-
Nenhum serviço de banco de dados separado necessário
-
Qdrant / pgvector
-
Postgres + BD vetorial
-
Várias
-
Nuvem gerenciada
-
Stack Docker (Core + Hub + Proxy)
-
Vector store
-
Oracle AI Database
-
Nenhuma
-
Nenhuma
-
-
-
Ciclo de vida da memória
-
Consolidação de 4 níveis + decaimento + auto-esquecimento
-
Extração passiva
-
Gerenciado pelo agente
-
Manual
-
Auto-esquecimento
-
Revisão manual; roteamento automático em andamento
-
Nenhum
-
Não informado
-
Decaimento + consolidação
-
Poda manual
-
-
-
Eficiência de tokens
-
~1,900 tokens/sessão ($10/ano)
-
Varia por integração
-
Memória principal no contexto
-
Varia
-
Precificação em nuvem
-
Não informado
-
Sem orçamento de tokens
-
Baseado em LLM (varia)
-
Varia
-
22K+ tokens em 240 observações
-
-
-
Viewer em tempo real
-
Sim (porta 3113)
-
Dashboard na nuvem
-
Dashboard na nuvem
-
Web UI
-
Dashboard na nuvem
-
Web UI do Hub
-
Não
-
Não
-
Não
-
Não
-
-
-
Self-hosted
-
Sim (padrão)
-
Opcional
-
Opcional
-
Sim
-
Não (somente nuvem)
-
Sim (Docker)
-
Sim
-
Sim (Oracle DB)
-
Sim
-
Sim
-
-
-
-Nota de benchmark: apenas o R@5 do agentmemory é resultado medido por nós mesmos (LongMemEval-S, reproduzível a partir de benchmark/COMPARISON.md). Os números de mem0 e Letta são os números LoCoMo publicados por eles (um dataset diferente); os números de MemPalace, supermemory, TencentDB (PersonaMem) e oracleagentmemory são alegações autorreportadas dos fornecedores que não reproduzimos de forma independente (a execução do oracleagentmemory usou GPT-5.5 contra um Oracle AI Database). Mostrados lado a lado apenas como ordem de grandeza, não como comparação direta sobre dados idênticos. As contagens de stars são aproximadas e mudam com o tempo.
-
-**Novos entrantes** que vale a pena conhecer, comparados em profundidade em [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Sistema | ⭐ | Abordagem |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Grafo de conhecimento temporal; os resultados publicados mais fortes em queries temporais (LongMemEval 63.8%), mas o grafo é construído de forma assíncrona, então fatos recentes podem atrasar |
-| Cognee | 30K | Ingestão de documento para grafo de conhecimento, somente Python, feito para extração estruturada de entidades em vez de captura de sessão |
+Comparação atualizada de produtos e opções de implantação, com fontes:
-O AgentMemory combina captura por agentes de programação, histórico de projeto pesquisável e visualizador local. O uso sem chave está disponível desde a primeira execução.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -777,7 +569,7 @@ Guia completo: [`integrations/hermes/`](../integrations/hermes/)
Inicie o servidor de memória: `npx -y @agentmemory/agentmemory@latest`
-#### Skills nativas via `npx skills add` (50+ agentes)
+#### Skills nativas via `npx skills add`
agentmemory entrega 17 skills no formato `/SKILL.md` estilo Claude Code: 9 skills de ação invocáveis (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) e 8 skills de referência que o agente carrega sob demanda (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). As skills de referência carregam tabelas de dados geradas a partir do código-fonte, então nunca ficam defasadas. O CLI [`skills`](https://npmjs.com/package/skills) da vercel-labs as instala automaticamente no diretório nativo de skills do agente chamador em 50+ agentes (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf e mais):
@@ -834,7 +626,7 @@ A entrada do agentmemory é o **mesmo bloco de servidor MCP** em todo host que u
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` instala o MCP e os hooks de captura no diretório de customização compartilhado. Veja [configuração e limites do Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` usa a mesma configuração de MCP e hooks das versões atuais do IDE. Instalações existentes devem atualizar com `--force`; veja as [notas de atualização](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` escreve a configuração no nível de usuário. Overrides de workspace vão em `.kiro/settings/mcp.json` ao lado do seu código. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` escreve o bloco `mcpServers` padrão. O Warp também auto-descobre skills de `.claude/skills/`; instalado o plugin do Claude Code, as 8 skills do agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) aparecem nativamente na paleta de slash commands do Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` escreve o bloco `mcpServers` padrão. O Warp também auto-descobre skills de `.claude/skills/`; instalado o plugin do Claude Code, as 9 skills do agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) aparecem nativamente na paleta de slash commands do Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` escreve o bloco `mcpServers` padrão. Usuários da extensão do VS Code: cole o mesmo bloco via Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (preferido) ou `config.json` (legado) | `agentmemory connect continue` cria `config.yaml` se estiver ausente ou modifica o `config.json` existente. Se o YAML já existir, mostra o bloco para adicionar em `mcpServers:` e preservar comentários e âncoras. O Continue usa um array, não um objeto. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` escreve sob `context_servers` (a chave do Zed, NÃO `mcpServers`). Servidores MCP remotos podem ser cabeados via `{"url": "..."}` em vez disso. |
@@ -931,7 +723,6 @@ Todo agente de codificação com IA vem com memória integrada: Claude Code tem
|---|---|---|
| Manutenção | Instruções mantidas manualmente | Histórico de sessões e memórias indexadas |
| Busca | Carrega tudo no contexto | BM25 + vector + graph (somente top-K) |
-| Custo em tokens | 22K+ em 240 observações | ~1,900 tokens (92% menos) |
| Entre agentes | Arquivos por agente | MCP + REST (qualquer agente) |
| Coordenação | Nenhuma | Leases, signals, actions, routines |
| Observabilidade | Leitura manual de arquivos | Viewer em tempo real na :3113 |
@@ -944,7 +735,7 @@ Todo agente de codificação com IA vem com memória integrada: Claude Code tem
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1029,7 +820,7 @@ Combinados com Reciprocal Rank Fusion (RRF, k=60) e diversificados por sessão (
Quando um índice vetorial está populado, o `mem::search` (por trás do `memory_recall`) usa o ranker híbrido BM25 + vector. Sem embeddings ele usa BM25. O `smart-search` pode ainda combinar correspondências estruturais do grafo quando existem dados de grafo, inclusive no modo sem chave. O recall de lessons roda sobre um índice BM25 dedicado em memória, em vez de escanear todo o corpus a cada query. Versões supersedidas de memória são excluídas de todo caminho de recall; a cadeia de versões mantém o histórico delas.
-Vetores sobrevivem a um crash ou force-kill. O índice vetorial é salvo em buckets no máximo a cada `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minutos). Todo vetor adicionado ou removido nesse meio tempo também é gravado imediatamente em um pequeno log pendente no state store, e o próximo início o reproduz sem chamar o provider de embedding. Cada save bem-sucedido esvazia o log. Documentos que ainda não têm vetor depois da reprodução são re-embeddados em segundo plano em lotes de `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) até que não reste nenhum, e um backfill interrompido continua no próximo início. `/agentmemory/status` e o viewer mostram o tamanho do log pendente e o estado do backfill. Instalações sem chave não escrevem nada.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
O BM25 tokeniza grego, cirílico, hebraico, árabe e latim acentuado nativamente. Para memórias em chinês / japonês / coreano, instale os segmentadores opcionais (`npm install @node-rs/jieba tiny-segmenter`) para dividir sequências CJK em tokens no nível de palavra; sem eles, o agentmemory cai graciosamente para tokenização da sequência inteira e imprime uma dica única no stderr.
@@ -1045,9 +836,9 @@ A instalação normal via npm inclui o runtime opcional `@huggingface/transforme
| Provider | Modelo | Custo | Notas |
|---|---|---|---|
-| **Local (opt-in recomendado)** | `all-MiniLM-L6-v2` | Gratuito | No dispositivo após o primeiro download do modelo, +8pp de recall sobre BM25 sozinho |
+| **Local (opt-in recomendado)** | `all-MiniLM-L6-v2` | Gratuito | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Camada gratuita | 100+ idiomas, dimensões 768/1536/3072 (MRL), entrada de 2048 tokens. Substitui o `text-embedding-004` ([descontinuado, desligamento em 14 de jan. de 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Maior qualidade |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Pago | Otimizado para código |
| Cohere | `embed-english-v3.0` | Trial gratuito | Uso geral |
| OpenRouter | Qualquer modelo | Varia | Proxy multi-modelo |
diff --git a/READMEs/README.pt-PT.md b/READMEs/README.pt-PT.md
index 1c280bcc21..94b2335b88 100644
--- a/READMEs/README.pt-PT.md
+++ b/READMEs/README.pt-PT.md
@@ -66,8 +66,6 @@
-
-
@@ -331,223 +329,17 @@ Notas da versão mais recente: [CHANGELOG.md](../CHANGELOG.md).
-
-
-
-
-### Precisão de Recuperação
-
-**coding-agent-life-v1** (corpus interno, reproduzível em sandbox)
-
-| Adaptador | P@5 | R@5 | Taxa de acerto top-5 | latência p50 |
-|---|---|---|---|---|
-| **agentmemory híbrido** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep de referência | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-Taxa de acerto top-5 de 100% no **teto matemático do P@5** para este corpus (0.240, ver scorecard). O híbrido recupera todas as sessões gold; o grep falha 1 de 2 golds na query temporal multi-sessão. O ganho está no **recall + temporal**, não na precisão agregada. Este benchmark é pequeno e escasso em golds; o LongMemEval-S maior, abaixo, diferencia melhor. Detalhamento completo por tipo + nota de correção: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 perguntas)
+Resultados atuais de recuperação, definições das métricas e passos de reprodução:
-| Sistema | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| fallback apenas BM25 | 86.2% | 94.6% | 71.5% |
-
-
-
-> Modelo de embedding: `all-MiniLM-L6-v2` (local, gratuito, sem chave de API). Relatórios completos: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Comparação com concorrentes: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) cobrindo agentmemory vs mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Reproduzir localmente:** [`eval/README.md`](../eval/README.md), um harness com adaptadores encaixáveis para o LongMemEval `_s` (500 perguntas públicas) + `coding-agent-life-v1` (corpus interno de 15 sessões). Os adaptadores grep / vetor / agentmemory são avaliados lado a lado, com saída NDJSON, e os scorecards publicados ficam em [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Combina-se com [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) e [Graphify](https://github.com/safishamsi/graphify).** Indexação de grafos de código, pipelines de build multiagente e grafos de conhecimento mais amplos em docs / PDFs / imagens / vídeos. O agentmemory recorda o trabalho; estes três projetos iluminam o resto da camada de contexto. Receitas + tabela de encaminhamento de perguntas: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Incorporado (CLAUDE.md)
-
-
-
Tipo
-
Motor de memória + servidor MCP
-
API de camada de memória
-
Runtime completo de agente
-
IA pessoal
-
API de memória + app
-
Hub de memória de equipa (proxy LLM)
-
Memória vetorial (OSS)
-
Motor de memória (Oracle DB)
-
Sistema de memória
-
Ficheiro estático
-
-
-
Recuperação R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Autorreportado
-
PersonaMem 76% (autorreportado)
-
~96.6% (autorreportado)
-
94.4% (autorreportado)
-
N/A
-
N/A (grep)
-
-
-
Captura automática
-
12 hooks (esforço manual zero)
-
Chamadas manuais a add()
-
Autoedição pelo agente
-
Manual
-
Extração do lado da API
-
Interceção via proxy (troca de URL base)
-
Manual
-
Extração via API
-
Manual
-
Edição manual
-
-
-
Pesquisa
-
BM25 + Vetor + Grafo (fusão RRF)
-
Vetor + Grafo
-
Vetor (arquivo)
-
Semântica
-
Vetor + RAG
-
4 tipos de recursos (Chat / Skill / Wiki / CodeGraph)
-
Apenas vetor
-
Vetor + semântica
-
Ponderada por decaimento
-
Carrega tudo para o contexto
-
-
-
Multiagente
-
MCP + REST + leases + signals
-
API (sem coordenação)
-
Apenas dentro do runtime da Letta
-
Não
-
Não
-
Papéis de equipa + recursos partilhados
-
Não
-
Apenas com âmbito definido
-
Partilhada multiagente
-
Ficheiros por agente
-
-
-
Dependência de framework
-
Nenhuma (qualquer cliente MCP)
-
Nenhuma
-
Alta (tem de usar Letta)
-
Autónomo
-
Nenhuma
-
O proxy intermedeia todas as chamadas ao modelo
-
Nenhuma
-
Oracle Database
-
Nenhuma
-
Formato por agente
-
-
-
Dependências externas
-
Nenhum serviço de base de dados separado necessário
-
Qdrant / pgvector
-
Postgres + BD vetorial
-
Várias
-
Nuvem gerida
-
Stack Docker (Core + Hub + Proxy)
-
Armazenamento vetorial
-
Oracle AI Database
-
Nenhuma
-
Nenhuma
-
-
-
Ciclo de vida da memória
-
Consolidação em 4 níveis + decaimento + esquecimento automático
-
Extração passiva
-
Gerido pelo agente
-
Manual
-
Esquecimento automático
-
Revisão manual; encaminhamento automático em desenvolvimento
-
Nenhum
-
Não indicado
-
Decaimento + consolidação
-
Limpeza manual
-
-
-
Eficiência de tokens
-
~1,900 tokens/sessão ($10/ano)
-
Varia com a integração
-
Memória principal no contexto
-
Varia
-
Preços na nuvem
-
Não indicado
-
Sem orçamento de tokens
-
Suportado por LLM (varia)
-
Varia
-
22K+ tokens em 240 observações
-
-
-
Visualizador em tempo real
-
Sim (porta 3113)
-
Painel na nuvem
-
Painel na nuvem
-
Interface Web
-
Painel na nuvem
-
Interface Web do hub
-
Não
-
Não
-
Não
-
Não
-
-
-
Auto-hospedado
-
Sim (predefinição)
-
Opcional
-
Opcional
-
Sim
-
Não (apenas na nuvem)
-
Sim (Docker)
-
Sim
-
Sim (Oracle DB)
-
Sim
-
Sim
-
-
-
-Nota sobre o benchmark: apenas o R@5 do agentmemory é um resultado medido por nós (LongMemEval-S, reproduzível a partir de benchmark/COMPARISON.md). Os valores de mem0 e Letta são os números LoCoMo publicados por eles (um dataset diferente); os valores de MemPalace, supermemory, TencentDB (PersonaMem) e oracleagentmemory são afirmações autorreportadas pelos fornecedores que não reproduzimos de forma independente (a execução do oracleagentmemory usou o GPT-5.5 contra uma Oracle AI Database). Mostrados lado a lado apenas como estimativa aproximada, não como comparação direta sobre dados idênticos. As contagens de estrelas são aproximadas e variam com o tempo.
-
-**Participantes mais recentes** que vale a pena conhecer, comparados em profundidade em [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Sistema | ⭐ | Abordagem |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Grafo de conhecimento temporal; os resultados publicados mais fortes em queries temporais (LongMemEval 63.8%), mas o grafo é construído de forma assíncrona, por isso os factos recentes podem atrasar |
-| Cognee | 30K | Ingestão de documentos para grafo de conhecimento, apenas Python, construído para extração estruturada de entidades em vez de captura de sessões |
+Comparação atualizada de produtos e opções de implementação, com fontes:
-O AgentMemory combina captura por agentes de programação, histórico de projeto pesquisável e visualizador local. O uso sem chave está disponível desde a primeira execução.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -779,7 +571,7 @@ Guia completo: [`integrations/hermes/`](../integrations/hermes/)
Inicie o servidor de memória: `npx -y @agentmemory/agentmemory@latest`
-#### Skills nativas via `npx skills add` (50+ agentes)
+#### Skills nativas via `npx skills add`
O agentmemory distribui 17 skills no formato `/SKILL.md`, no estilo do Claude Code: 9 skills de ação invocáveis (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) e 8 skills de referência que o agente carrega sob demanda (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). As skills de referência trazem tabelas de dados geradas a partir do código-fonte, por isso nunca ficam desatualizadas. O CLI [`skills`](https://npmjs.com/package/skills) da vercel-labs instala-as automaticamente no diretório nativo de skills do agente que o invoca, em mais de 50 agentes (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, e outros):
@@ -836,7 +628,7 @@ A entrada do agentmemory é o **mesmo bloco de servidor MCP** em todos os anfitr
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` instala o MCP e os hooks de captura no diretório de personalização partilhado. Consulte [configuração e limites do Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` usa a mesma configuração de MCP e hooks das versões atuais do IDE. As instalações existentes devem atualizar com `--force`; consulte as [notas de atualização](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` escreve a configuração ao nível do utilizador. As substituições ao nível do workspace vão em `.kiro/settings/mcp.json`, junto ao seu código. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` escreve o bloco `mcpServers` padrão. O Warp também deteta automaticamente skills em `.claude/skills/`; depois de instalado o plugin do Claude Code, as 8 skills do agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) aparecem nativamente na paleta de comandos slash do Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` escreve o bloco `mcpServers` padrão. O Warp também deteta automaticamente skills em `.claude/skills/`; depois de instalado o plugin do Claude Code, as 9 skills do agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) aparecem nativamente na paleta de comandos slash do Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` escreve o bloco `mcpServers` padrão. Utilizadores da extensão VS Code: cole o mesmo bloco através de Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (preferido) ou `config.json` (legado) | `agentmemory connect continue` cria `config.yaml` se estiver ausente ou modifica o `config.json` existente. Se o YAML já existir, mostra o bloco para adicionar em `mcpServers:` e preservar comentários e âncoras. O Continue usa um array, não um objeto. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` escreve sob `context_servers` (a chave do Zed, NÃO `mcpServers`). Servidores MCP remotos podem, em alternativa, ser ligados através de `{"url": "..."}`. |
@@ -927,7 +719,6 @@ Todo o agente de codificação com IA vem com memória incorporada: o Claude Cod
|---|---|---|
| Manutenção | Instruções mantidas manualmente | Histórico de sessões e memórias indexadas |
| Pesquisa | Carrega tudo para o contexto | BM25 + vetor + grafo (apenas top-K) |
-| Custo em tokens | 22K+ em 240 observações | ~1,900 tokens (92% menos) |
| Entre agentes | Ficheiros por agente | MCP + REST (qualquer agente) |
| Coordenação | Nenhuma | Leases, signals, actions, routines |
| Observabilidade | Ler ficheiros manualmente | Visualizador em tempo real na :3113 |
@@ -940,7 +731,7 @@ Todo o agente de codificação com IA vem com memória incorporada: o Claude Cod
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1025,7 +816,7 @@ Combinados com Reciprocal Rank Fusion (RRF, k=60) e diversificados por sessão (
Quando um índice vetorial está preenchido, o `mem::search` (por trás de `memory_recall`) usa o ranker híbrido BM25 + vetor. Sem embeddings, usa o BM25. O `smart-search` pode ainda combinar correspondências estruturais do grafo quando existem dados de grafo, inclusive em modo sem chave. O recall de lessons corre sobre um índice BM25 dedicado em memória, em vez de percorrer todo o corpus a cada query. As versões de memória substituídas são excluídas de todos os caminhos de recall; a cadeia de versões mantém o seu histórico.
-Os vetores sobrevivem a um crash ou a uma paragem forçada. O índice vetorial é guardado em blocos no máximo a cada `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minutos). Cada vetor adicionado ou removido entretanto também é escrito imediatamente num pequeno registo pendente no state store, e o arranque seguinte reproduz esse registo sem chamar o fornecedor de embeddings. Cada gravação bem-sucedida esvazia o registo. Os documentos que ainda não têm vetor depois da reprodução são re-embutidos (re-embedded) em segundo plano, em lotes de `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500), até não restar nenhum, e um preenchimento retroativo (backfill) interrompido continua no arranque seguinte. O `/agentmemory/status` e o visualizador mostram o tamanho do registo pendente e o estado do backfill. As instalações sem chave não escrevem nada.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
O BM25 tokeniza grego, cirílico, hebraico, árabe e latim acentuado de fábrica. Para memórias em chinês / japonês / coreano, instale os segmentadores opcionais (`npm install @node-rs/jieba tiny-segmenter`) para dividir sequências CJK em tokens ao nível da palavra; sem eles, o agentmemory recua suavemente para a tokenização da sequência completa e imprime uma dica única no stderr.
@@ -1041,9 +832,9 @@ A instalação normal via npm inclui o runtime opcional `@huggingface/transforme
| Fornecedor | Modelo | Custo | Notas |
|---|---|---|---|
-| **Local (opt-in recomendado)** | `all-MiniLM-L6-v2` | Gratuito | No dispositivo após o primeiro download do modelo, +8pp de recall sobre apenas BM25 |
+| **Local (opt-in recomendado)** | `all-MiniLM-L6-v2` | Gratuito | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Nível gratuito | 100+ idiomas, dimensões 768/1536/3072 (MRL), entrada de 2048 tokens. Substitui o `text-embedding-004` ([descontinuado, encerramento a 14 de janeiro de 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Qualidade mais alta |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Pago | Otimizado para código |
| Cohere | `embed-english-v3.0` | Teste gratuito | Uso geral |
| OpenRouter | Qualquer modelo | Varia | Proxy multimodelo |
diff --git a/READMEs/README.ro-RO.md b/READMEs/README.ro-RO.md
index b2401fe6b1..1cefbf8672 100644
--- a/READMEs/README.ro-RO.md
+++ b/READMEs/README.ro-RO.md
@@ -66,8 +66,6 @@
-
-
@@ -333,223 +331,17 @@ Notele celei mai recente versiuni: [CHANGELOG.md](../CHANGELOG.md).
-
-
-
-
-### Acuratețea regăsirii
-
-**coding-agent-life-v1** (corpus intern, reproductibil în sandbox)
-
-| Adaptor | P@5 | R@5 | Rata de succes Top-5 | Latență p50 |
-|---|---|---|---|---|
-| **agentmemory hibrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep (referință) | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-Rată de succes Top-5 de 100% la **plafonul matematic P@5** pentru acest corpus (0.240, vezi scorecard-ul). Modul hibrid regăsește fiecare sesiune de referință (gold); grep ratează 1 din 2 sesiuni de referință la interogarea temporală pe mai multe sesiuni. Avantajul este pe **recall + temporal**, nu pe precizia agregată. Acest benchmark este mic și are puține exemple de referință; LongMemEval-S, mai amplu, de mai jos diferențiază mai bine. Detalierea completă pe tipuri + nota de corecție: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 de întrebări)
+Rezultatele actuale ale căutării, definițiile metricilor și pașii de reproducere:
-| Sistem | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| Doar BM25 (fallback) | 86.2% | 94.6% | 71.5% |
-
-
-
-> Model de embedding: `all-MiniLM-L6-v2` (local, gratuit, fără cheie API). Rapoarte complete: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Comparație cu competitorii: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md), care acoperă agentmemory vs mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Reproduce local:** [`eval/README.md`](../eval/README.md), un harness cu adaptoare interschimbabile pentru LongMemEval `_s` (500 de întrebări publice) + `coding-agent-life-v1` (corpus intern de 15 sesiuni). Adaptoarele grep / vector / agentmemory sunt scorate în paralel, cu output NDJSON; scorecard-urile publicate ajung în [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Se combină bine cu [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) și [Graphify](https://github.com/safishamsi/graphify).** Indexare code-graph, pipeline-uri de build multi-agent și grafuri de cunoștințe mai ample pe documente / PDF-uri / imagini / videoclipuri. agentmemory reține munca; aceste trei proiecte iluminează restul stratului de context. Rețete + tabel de direcționare a întrebărilor: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Încorporat (CLAUDE.md)
-
-
-
Tip
-
Motor de memorie + server MCP
-
API strat de memorie
-
Runtime complet de agent
-
AI personal
-
API de memorie + aplicație
-
Hub de memorie de echipă (proxy LLM)
-
Memorie vectorială (OSS)
-
Motor de memorie (Oracle DB)
-
Sistem de memorie
-
Fișier static
-
-
-
Regăsire R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Auto-raportat
-
PersonaMem 76% (auto-raportat)
-
~96.6% (auto-raportat)
-
94.4% (auto-raportat)
-
N/A
-
N/A (grep)
-
-
-
Captare automată
-
12 hook-uri (fără efort manual)
-
Apeluri manuale add()
-
Auto-editare de către agent
-
Manual
-
Extracție pe partea de API
-
Interceptare prin proxy (schimbare base-URL)
-
Manual
-
Extracție API
-
Manual
-
Editare manuală
-
-
-
Căutare
-
BM25 + Vector + Graf (fuziune RRF)
-
Vector + Graf
-
Vector (arhivă)
-
Semantică
-
Vector + RAG
-
4 tipuri de active (Chat / Skill / Wiki / CodeGraph)
-
Doar vector
-
Vector + semantică
-
Ponderată prin declin (decay)
-
Încarcă totul în context
-
-
-
Multi-agent
-
MCP + REST + lease-uri + semnale
-
API (fără coordonare)
-
Doar în interiorul runtime-ului Letta
-
Nu
-
Nu
-
Roluri de echipă + active partajate
-
Nu
-
Doar cu scope limitat
-
Partajat multi-agent
-
Fișiere per agent
-
-
-
Dependență de framework
-
Niciuna (orice client MCP)
-
Niciuna
-
Ridicată (trebuie folosit Letta)
-
Independent (standalone)
-
Niciuna
-
Proxy-ul intermediază fiecare apel de model
-
Niciuna
-
Oracle Database
-
Niciuna
-
Format per agent
-
-
-
Dependențe externe
-
Nu necesită un serviciu de baze de date separat
-
Qdrant / pgvector
-
Postgres + bază de date vectorială
-
Multiple
-
Cloud administrat
-
Stivă Docker (Core + Hub + Proxy)
-
Magazin vectorial
-
Oracle AI Database
-
Niciuna
-
Niciuna
-
-
-
Ciclul de viață al memoriei
-
Consolidare pe 4 niveluri + declin + uitare automată
-
Extracție pasivă
-
Gestionat de agent
-
Manual
-
Uitare automată
-
Revizuire manuală; direcționare automată în lucru
-
Niciuna
-
Nespecificat
-
Declin + consolidare
-
Curățare manuală
-
-
-
Eficiența tokenilor
-
~1,900 tokeni/sesiune ($10/an)
-
Variază în funcție de integrare
-
Memorie de bază (core) în context
-
Variază
-
Preț cloud
-
Nespecificat
-
Fără buget de tokeni
-
Susținut de LLM (variază)
-
Variază
-
22K+ tokeni la 240 observații
-
-
-
Vizualizator în timp real
-
Da (port 3113)
-
Dashboard cloud
-
Dashboard cloud
-
Interfață web
-
Dashboard cloud
-
Interfață web Hub
-
Nu
-
Nu
-
Nu
-
Nu
-
-
-
Auto-hostat
-
Da (implicit)
-
Opțional
-
Opțional
-
Da
-
Nu (doar cloud)
-
Da (Docker)
-
Da
-
Da (Oracle DB)
-
Da
-
Da
-
-
-
-Notă despre benchmark: doar R@5 pentru agentmemory este un rezultat măsurat de noi înșine (LongMemEval-S, reproductibil din benchmark/COMPARISON.md). Cifrele pentru mem0 și Letta sunt numerele lor publicate pe LoCoMo (un set de date diferit); cifrele pentru MemPalace, supermemory, TencentDB (PersonaMem) și oracleagentmemory sunt afirmații auto-raportate de furnizori, pe care nu le-am reprodus independent (rularea oracleagentmemory a folosit GPT-5.5 împotriva unei Oracle AI Database). Prezentate una lângă alta doar orientativ, nu ca o comparație directă pe date identice. Numărul de stele este aproximativ și variază în timp.
-
-**Jucători mai noi** merită cunoscuți, comparați în detaliu în [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Sistem | ⭐ | Unghi |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Graf de cunoștințe temporal; cele mai bune rezultate publicate pentru interogări temporale (LongMemEval 63.8%), dar graful se construiește asincron, așa că faptele noi pot întârzia |
-| Cognee | 30K | Ingestie document-în-graf-de-cunoștințe, doar Python, construit pentru extracție structurată de entități, nu pentru captarea sesiunilor |
+Comparație actualizată a produselor și opțiunilor de implementare, cu surse:
-AgentMemory combină captura sesiunilor, istoricul căutabil al proiectului și un viewer local, cu funcționare fără cheie de la prima pornire.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -786,7 +578,7 @@ Ghid complet: [`integrations/hermes/`](../integrations/hermes/)
Pornește serverul de memorie: `npx -y @agentmemory/agentmemory@latest`
-#### Skill-uri native prin `npx skills add` (50+ agenți)
+#### Skill-uri native prin `npx skills add`
agentmemory vine cu 17 skill-uri în formatul `/SKILL.md`, în stilul Claude Code: 9 skill-uri de acțiune invocabile (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) și 8 skill-uri de referință pe care agentul le încarcă la cerere (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Skill-urile de referință conțin tabele de date generate din sursă, așa că nu devin niciodată desincronizate. CLI-ul [`skills`](https://npmjs.com/package/skills) de la vercel-labs le instalează automat în directorul nativ de skill-uri al agentului apelant, pe 50+ agenți (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf și altele):
@@ -844,7 +636,7 @@ Intrarea agentmemory este **același bloc de server MCP** pe fiecare host care f
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` instalează MCP și hook-urile de captare în directorul partajat de personalizare. Vezi [configurarea și limitele Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` folosește aceeași configurație MCP și de hook-uri ca versiunile IDE actuale. Instalările existente ar trebui să se reîmprospăteze cu `--force`; vezi [notele de actualizare](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` scrie configurația la nivel de utilizator. Suprascrierile la nivel de workspace se pun în `.kiro/settings/mcp.json`, alături de codul tău. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` scrie blocul standard `mcpServers`. Warp detectează automat și skill-urile din `.claude/skills/`; odată instalat plugin-ul Claude Code, cele 8 skill-uri agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) apar nativ în paleta de comenzi slash a Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` scrie blocul standard `mcpServers`. Warp detectează automat și skill-urile din `.claude/skills/`; odată instalat plugin-ul Claude Code, cele 9 skill-uri agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) apar nativ în paleta de comenzi slash a Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` scrie blocul standard `mcpServers`. Utilizatorii extensiei VS Code: lipește același bloc prin Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (preferat) sau `config.json` (vechi/legacy) | `agentmemory connect continue` creează `config.yaml` când lipsește configurarea sau modifică `config.json` existent. Pentru YAML existent afișează blocul de adăugat la `mcpServers:`, păstrând comentariile și ancorele. Continue folosește o listă, nu un obiect. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` scrie sub `context_servers` (cheia specifică Zed, NU `mcpServers`). Serverele MCP la distanță pot fi conectate, în schimb, prin `{"url": "..."}`. |
@@ -921,7 +713,6 @@ Fiecare agent de programare AI vine cu memorie încorporată: Claude Code are `M
|---|---|---|
| Întreținere | Instrucțiuni administrate manual | Istoric de sesiuni și memorie indexate |
| Căutare | Încarcă totul în context | BM25 + vector + graf (doar top-K) |
-| Cost în tokeni | 22K+ la 240 de observații | ~1,900 tokeni (92% mai puțin) |
| Între agenți | Fișiere per agent | MCP + REST (orice agent) |
| Coordonare | Niciuna | Lease-uri, semnale, acțiuni, rutine |
| Observabilitate | Citire manuală a fișierelor | Vizualizator în timp real pe :3113 |
@@ -934,7 +725,7 @@ Fiecare agent de programare AI vine cu memorie încorporată: Claude Code are `M
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1019,7 +810,7 @@ Fuzionate prin Reciprocal Rank Fusion (RRF, k=60) și diversificate pe sesiune (
Când indexul vectorial este populat, `mem::search` (din spatele lui `memory_recall`) folosește clasificatorul hibrid BM25 + vector. Fără embeddings, folosește BM25. `smart-search` poate fuziona, în plus, potriviri structurale din graf, atunci când există date de graf, inclusiv în modul keyless. Reamintirea lecțiilor rulează pe un index BM25 dedicat, în memorie, în loc să scaneze întregul corpus la fiecare interogare. Versiunile de memorie suprascrise sunt excluse din fiecare rută de reamintire; lanțul de versiuni le păstrează istoricul.
-Vectorii supraviețuiesc unui crash sau unei terminări forțate (force-kill). Indexul vectorial este salvat pe loturi (buckets) cel mult o dată la fiecare `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minute). Fiecare vector adăugat sau eliminat între timp este scris imediat și într-un mic jurnal de așteptare (pending log) din state store, iar următoarea pornire îl reaplică fără a apela furnizorul de embedding. Fiecare salvare reușită golește jurnalul. Documentele care încă nu au un vector după reaplicare sunt reîncorporate (re-embedded) în fundal, pe loturi de `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500), până când nu mai rămâne niciunul, iar o reumplere (backfill) întreruptă continuă la următoarea pornire. `/agentmemory/status` și vizualizatorul arată dimensiunea jurnalului de așteptare și starea reumplerii. Instalările keyless nu scriu nimic.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokenizează, din oficiu, greaca, chirilica, ebraica, araba și latina cu diacritice. Pentru memorii în chineză / japoneză / coreeană, instalează segmentatoarele opționale (`npm install @node-rs/jieba tiny-segmenter`) pentru a împărți secvențele CJK în tokeni la nivel de cuvânt; fără ele, agentmemory trece ușor la tokenizarea întregii secvențe și afișează o singură dată un indiciu pe stderr.
@@ -1035,9 +826,9 @@ Instalarea npm normală include runtime-ul opțional `@huggingface/transformers`
| Furnizor | Model | Cost | Note |
|---|---|---|---|
-| **Local (opțiune recomandată)** | `all-MiniLM-L6-v2` | Gratuit | Pe dispozitiv, după prima descărcare a modelului; +8pp recall față de doar-BM25 |
+| **Local (opțiune recomandată)** | `all-MiniLM-L6-v2` | Gratuit | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Nivel gratuit | 100+ limbi, 768/1536/3072 dimensiuni (MRL), input de 2048 de tokeni. Înlocuiește `text-embedding-004` ([depreciat, dezactivat 14 ian. 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Cea mai bună calitate |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Plătit | Optimizat pentru cod |
| Cohere | `embed-english-v3.0` | Perioadă de probă gratuită | Scop general |
| OpenRouter | Orice model | Variabil | Proxy multi-model |
diff --git a/READMEs/README.ru-RU.md b/READMEs/README.ru-RU.md
index 83bbaf21e2..33558b5880 100644
--- a/READMEs/README.ru-RU.md
+++ b/READMEs/README.ru-RU.md
@@ -66,8 +66,6 @@
-
-Заметка о бенчмарках: только R@5 agentmemory — наш собственный замер (LongMemEval-S, воспроизводимо из benchmark/COMPARISON.md). Цифры mem0 и Letta — их опубликованные результаты LoCoMo (другой датасет); цифры MemPalace, supermemory, TencentDB (PersonaMem) и oracleagentmemory — самозаявленные вендорами значения, которые мы независимо не воспроизводили (прогон oracleagentmemory использовал GPT-5.5 против Oracle AI Database). Показаны рядом только для ориентира, это не сравнение лоб в лоб на одинаковых данных. Количество звёзд приблизительно и меняется со временем.
-
-**Более новые участники**, которых стоит знать, подробно сравниваются в [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Система | ⭐ | Особенность |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Темпоральный граф знаний; сильнейшие опубликованные результаты на темпоральных запросах (LongMemEval 63.8%), но граф строится асинхронно, поэтому свежие факты могут запаздывать |
-| Cognee | 30K | Превращение документов в граф знаний, только Python, создан для структурированного извлечения сущностей, а не для захвата сессий |
+Актуальное сравнение продуктов и вариантов развёртывания с источниками:
-AgentMemory объединяет захват сессий, поиск по истории проекта и локальный viewer; работа без ключа доступна с первого запуска.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -786,7 +578,7 @@ Verify with `curl http://localhost:3111/agentmemory/health`. Open http://localho
Запустите сервер памяти: `npx -y @agentmemory/agentmemory@latest`
-#### Нативные skill'ы через `npx skills add` (50+ агентов)
+#### Нативные skill'ы через `npx skills add`
agentmemory поставляет 17 skill'ов в формате `/SKILL.md` в стиле Claude Code: 9 вызываемых action-skill'ов (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) и 8 справочных skill'ов, которые агент подгружает по мере надобности (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Справочные skill'ы содержат таблицы данных, сгенерированные из исходников, поэтому они никогда не устаревают. CLI [`skills`](https://npmjs.com/package/skills) от vercel-labs автоматически устанавливает их в нативный каталог skill'ов вызывающего агента для 50+ агентов (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf и другие):
@@ -844,7 +636,7 @@ npx skills add rohitg00/agentmemory -y -a '*' # install to every installed age
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` устанавливает MCP и хуки захвата в общем каталоге настроек. См. [установку и ограничения Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` использует ту же конфигурацию MCP и хуков, что и текущие версии IDE. Существующие установки следует обновить с `--force`; см. [заметки об обновлении](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` записывает конфиг на уровне пользователя. Переопределения на уровне workspace — в `.kiro/settings/mcp.json` рядом с кодом. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` записывает стандартный блок `mcpServers`. Warp также автоматически обнаруживает skill'ы из `.claude/skills/`; как только установлен плагин Claude Code, 8 skill'ов agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) нативно появляются в палитре slash-команд Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` записывает стандартный блок `mcpServers`. Warp также автоматически обнаруживает skill'ы из `.claude/skills/`; как только установлен плагин Claude Code, 9 skill'ов agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) нативно появляются в палитре slash-команд Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` записывает стандартный блок `mcpServers`. Пользователи расширения VS Code: вставьте тот же блок через Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (предпочтительно) или `config.json` (legacy) | `agentmemory connect continue` создаёт `config.yaml`, если настройки отсутствуют, или меняет существующий `config.json`. Для существующего YAML выводит блок для `mcpServers:`, сохраняя комментарии и якоря. Continue использует массив, а не объект. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` пишет под `context_servers` (ключ Zed, НЕ `mcpServers`). Удалённые MCP-серверы можно вместо этого подключить через `{"url": "..."}`. |
@@ -927,7 +719,6 @@ REST использует `3111`, viewer `3113`. Оставьте их в при
|---|---|---|
| Обслуживание | Инструкции, редактируемые вручную | Индексируемая история сессий и воспоминания |
| Поиск | Загружает всё в контекст | BM25 + вектор + граф (только top-K) |
-| Цена в токенах | 22K+ при 240 наблюдениях | ~1,900 токенов (на 92% меньше) |
| Между агентами | Файлы на каждого агента | MCP + REST (любой агент) |
| Координация | Нет | Lease'ы, сигналы, action'ы, routine'ы |
| Наблюдаемость | Читать файлы вручную | Просмотрщик в реальном времени на :3113 |
@@ -940,7 +731,7 @@ REST использует `3111`, viewer `3113`. Оставьте их в при
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1025,7 +816,7 @@ SessionStart hook fires
Когда векторный индекс заполнен, `mem::search` (за которым стоит `memory_recall`) использует гибридный ранжировщик BM25 + вектор. Без эмбеддингов используется BM25. `smart-search` может дополнительно сливать структурные совпадения из графа, если граф-данные существуют, в том числе в режиме без ключей. Recall уроков работает на выделенном in-memory BM25-индексе вместо сканирования всего корпуса на каждый запрос. Вытесненные (superseded) версии памяти исключаются из каждого пути recall; цепочка версий хранит их историю.
-Векторы переживают крэш или force-kill. Векторный индекс сохраняется бакетами не чаще, чем раз в `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 минут). Каждый добавленный или удалённый в промежутке вектор также сразу записывается в небольшой отложенный лог в state store, и при следующем запуске он воспроизводится без обращения к провайдеру эмбеддингов. Каждое успешное сохранение очищает лог. Документы, у которых после воспроизведения всё ещё нет вектора, переэмбеддятся в фоне партиями по `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500), пока не останется ни одного, а остановленный backfill продолжается при следующем запуске. `/agentmemory/status` и просмотрщик показывают размер отложенного лога и состояние backfill'а. Установки без ключей ничего не пишут.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 «из коробки» токенизирует греческий, кириллицу, иврит, арабский и латиницу с диакритикой. Для записей на китайском / японском / корейском поставьте опциональные сегментаторы (`npm install @node-rs/jieba tiny-segmenter`), чтобы разбивать CJK-последовательности на токены уровня слова; без них agentmemory мягко откатывается к токенизации целых последовательностей и один раз выводит подсказку в stderr.
@@ -1041,9 +832,9 @@ EMBEDDING_PROVIDER=local
| Провайдер | Модель | Стоимость | Заметки |
|---|---|---|---|
-| **Локально (рекомендуется для подключения)** | `all-MiniLM-L6-v2` | Бесплатно | На устройстве после первой загрузки модели, +8 пп recall по сравнению только с BM25 |
+| **Локально (рекомендуется для подключения)** | `all-MiniLM-L6-v2` | Бесплатно | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Бесплатный тариф | 100+ языков, размерности 768/1536/3072 (MRL), вход 2048 токенов. Заменяет `text-embedding-004` ([устарел, отключение 14 янв. 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Высочайшее качество |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Платно | Оптимизирован под код |
| Cohere | `embed-english-v3.0` | Бесплатная пробная версия | Общего назначения |
| OpenRouter | Любая модель | Зависит | Мульти-модельный прокси |
diff --git a/READMEs/README.sv-SE.md b/READMEs/README.sv-SE.md
index 3aa899882f..760ca65053 100644
--- a/READMEs/README.sv-SE.md
+++ b/READMEs/README.sv-SE.md
@@ -66,8 +66,6 @@
-
-
@@ -331,223 +329,17 @@ Senaste versionsanteckningarna: [CHANGELOG.md](../CHANGELOG.md).
-
-
-
-
-### Hämtningsnoggrannhet
-
-**coding-agent-life-v1** (internt korpus, reproducerbart i sandbox)
-
-| Adapter | P@5 | R@5 | Topp-5-träffrate | p50-latens |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep-baslinje | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-100% topp-5-träffrate vid **P@5-mattaket** för detta korpus (0.240, se scorecard). Hybrid hämtar varje gold-session; grep missar 1 av 2 gold på den flersessions-temporala frågan. Lyftet ligger i **recall + temporal**, inte i aggregerad precision. Detta benchmark är litet och gold-glest; det större LongMemEval-S nedan differentierar bättre. Fullständig uppdelning per typ + rättelsenotis: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 frågor)
+Aktuella sökresultat, definitioner av mått och steg för reproduktion:
-| System | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| Endast BM25-fallback | 86.2% | 94.6% | 71.5% |
-
-
-
-Benchmark-notis: det är bara agentmemorys R@5 som är vårt eget uppmätta resultat (LongMemEval-S, reproducerbart från benchmark/COMPARISON.md). Siffrorna för mem0 och Letta är deras publicerade LoCoMo-siffror (ett annat dataset); siffrorna för MemPalace, supermemory, TencentDB (PersonaMem) och oracleagentmemory är leverantörens egna rapporterade påståenden som vi inte har reproducerat oberoende (oracleagentmemorys körning använde GPT-5.5 mot en Oracle AI Database). Visas sida vid sida enbart som fingervisning, inte en head-to-head på identisk data. Stjärnantal är ungefärliga och förändras över tid.
-
-**Nyare aktörer** värda att känna till, jämförda på djupet i [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| System | ⭐ | Inriktning |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporal kunskapsgraf; starkast publicerade resultat för temporala frågor (LongMemEval 63.8%), men grafen byggs asynkront så färska fakta kan ligga efter |
-| Cognee | 30K | Dokument-till-kunskapsgraf-inmatning, endast Python, byggt för strukturerad entitetsextraktion snarare än sessionsinfångning |
+Uppdaterad jämförelse av produkter och driftsalternativ med källor:
-AgentMemory kombinerar registrering från kodningsagenter, sökbar projekthistorik och en lokal viewer. Du kan arbeta utan nyckel från första starten.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -779,7 +571,7 @@ Fullständig guide: [`integrations/hermes/`](../integrations/hermes/)
Starta minnesservern: `npx -y @agentmemory/agentmemory@latest`
-#### Inbyggda skills via `npx skills add` (50+ agenter)
+#### Inbyggda skills via `npx skills add`
agentmemory levereras med 17 skills i Claude-Code-stilens `/SKILL.md`-format: 9 anropsbara handlings-skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) och 8 referens-skills som agenten laddar vid behov (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Referens-skillsen innehåller datatabeller genererade från källkoden, så de avviker aldrig. CLI-verktyget [`skills`](https://npmjs.com/package/skills) från vercel-labs installerar dem automatiskt i den anropande agentens inbyggda skill-katalog över 50+ agenter (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf och fler):
@@ -836,7 +628,7 @@ agentmemory-posten är **samma MCP-serverblock** på alla värdar som använder
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` installerar MCP och infångnings-hooks i den delade anpassningskatalogen. Se [Antigravity-konfiguration och begränsningar](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` använder samma MCP- och hook-konfiguration som nuvarande IDE-versioner. Befintliga installationer bör uppdateras med `--force`; se [uppgraderingsanteckningarna](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` skriver konfigurationen på användarnivå. Arbetsyteöverskrivningar läggs i `.kiro/settings/mcp.json` bredvid din kod. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` skriver det vanliga `mcpServers`-blocket. Warp upptäcker också automatiskt skills från `.claude/skills/`; när Claude Code-pluginet är installerat visas agentmemorys 8 skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) naturligt i Warps snabbkommando-palett. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` skriver det vanliga `mcpServers`-blocket. Warp upptäcker också automatiskt skills från `.claude/skills/`; när Claude Code-pluginet är installerat visas agentmemorys 9 skills (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) naturligt i Warps snabbkommando-palett. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` skriver det vanliga `mcpServers`-blocket. VS Code-tilläggsanvändare: klistra in samma block via Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (att föredra) eller `config.json` (äldre) | `agentmemory connect continue` skapar en saknad `config.yaml` eller ändrar befintlig `config.json`. Om YAML-filen finns visas blocket för `mcpServers:` så att kommentarer och ankare bevaras. Continue använder en array, inte ett objekt. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` skriver under `context_servers` (Zeds nyckel, INTE `mcpServers`). Fjärranslutna MCP-servrar kan istället kopplas in via `{"url": "..."}`. |
@@ -933,7 +725,6 @@ Varje AI-kodningsagent levereras med inbyggt minne: Claude Code har `MEMORY.md`,
|---|---|---|
| Underhåll | Manuellt underhållna instruktioner | Indexerad sessionshistorik och minnen |
| Sökning | Laddar allt in i kontexten | BM25 + vektor + graf (endast topp-K) |
-| Tokenkostnad | 22K+ vid 240 observationer | ~1,900 tokens (92% mindre) |
| Mellan agenter | Filer per agent | MCP + REST (vilken agent som helst) |
| Koordinering | Ingen | Leaser, signaler, åtgärder, rutiner |
| Observerbarhet | Läs filer manuellt | Viewer i realtid på :3113 |
@@ -946,7 +737,7 @@ Varje AI-kodningsagent levereras med inbyggt minne: Claude Code har `MEMORY.md`,
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1031,7 +822,7 @@ Sammanfogad med Reciprocal Rank Fusion (RRF, k=60) och sessionsdiversifierad (ma
När ett vektorindex är ifyllt använder `mem::search` (bakom `memory_recall`) den hybrida BM25 + vektor-rankaren. Utan embeddings används BM25. `smart-search` kan dessutom sammanfoga strukturella grafmatchningar när grafdata finns, även i nyckelfritt läge. Återkallning av lärdomar körs på ett dedikerat BM25-index i minnet istället för att skanna hela korpuset per fråga. Ersatta minnesversioner exkluderas från varje återkallningsväg; versionskedjan behåller deras historik.
-Vektorer överlever en krasch eller en tvingad avstängning. Vektorindexet sparas i buckets högst var `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minuter). Varje vektor som läggs till eller tas bort däremellan skrivs också direkt till en liten väntande logg i state-lagret, och nästa start spelar upp den igen utan att anropa embedding-leverantören. Varje lyckad sparning tömmer loggen. Dokument som fortfarande saknar en vektor efter uppspelningen bäddas in igen i bakgrunden i omgångar om `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) tills inga återstår, och en backfill som stoppas fortsätter vid nästa start. `/agentmemory/status` och viewern visar storleken på den väntande loggen och backfill-tillståndet. Nyckelfria installationer skriver ingenting.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokeniserar grekiska, kyrilliska, hebreiska, arabiska och accentuerad latin direkt ur lådan. För kinesiska/japanska/koreanska minnen, installera de valfria segmenterarna (`npm install @node-rs/jieba tiny-segmenter`) för att dela upp CJK-sekvenser i ordnivå-tokens; utan dem faller agentmemory mjukt tillbaka till tokenisering av hela sekvenser och skriver ut en engångstips på stderr.
@@ -1047,9 +838,9 @@ Den vanliga npm-installationen inkluderar den valfria körtiden `@huggingface/tr
| Leverantör | Modell | Kostnad | Anteckningar |
|---|---|---|---|
-| **Lokal (rekommenderat att aktivera)** | `all-MiniLM-L6-v2` | Kostnadsfri | På enheten efter den första modellnedladdningen, +8 procentenheter recall jämfört med endast BM25 |
+| **Lokal (rekommenderat att aktivera)** | `all-MiniLM-L6-v2` | Kostnadsfri | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Kostnadsfri nivå | 100+ språk, 768/1536/3072 dimensioner (MRL), 2048-token indata. Ersätter `text-embedding-004` ([utfasad, stängs ner 14 jan 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Högst kvalitet |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Betald | Optimerad för kod |
| Cohere | `embed-english-v3.0` | Kostnadsfri provperiod | Generell användning |
| OpenRouter | Vilken modell som helst | Varierar | Multimodell-proxy |
diff --git a/READMEs/README.th-TH.md b/READMEs/README.th-TH.md
index 93b2dde01c..d67522366c 100644
--- a/READMEs/README.th-TH.md
+++ b/READMEs/README.th-TH.md
@@ -66,8 +66,6 @@
-
-### Katumpakan ng Retrieval
-
-**coding-agent-life-v1** (in-house corpus, reproducible sa sandbox)
-
-| Adapter | P@5 | R@5 | Top-5 hit rate | p50 latency |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep baseline | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-100% top-5 hit rate sa **P@5 math ceiling** para sa corpus na ito (0.240, tingnan ang scorecard). Nakukuha ng hybrid ang bawat gold session; nawawala sa grep ang 1 sa 2 gold sa multi-session temporal query. Ang lift ay sa **recall + temporal**, hindi sa aggregate precision. Maliit at gold-sparse ang benchmark na ito; mas mahusay na nagkakaiba ang mas malaking LongMemEval-S sa ibaba. Buong per-type breakdown + correction note: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 tanong)
+Kasalukuyang resulta ng paghahanap, kahulugan ng mga sukatan, at mga hakbang sa pag-ulit ng pagsubok:
-| Sistema | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| BM25-only fallback | 86.2% | 94.6% | 71.5% |
-
-
-
-
-### Pagtitipid sa Token
-
-| Pamamaraan | Tokens/yr | Cost/yr |
-|---|---|---|
-| I-paste ang buong context | 19.5M+ | Imposible (lumalagpas sa window) |
-| Pinaikli ng LLM | ~650K | ~$500 |
-| **agentmemory** | **~170K** | **~$10** |
-| agentmemory + local embeddings | ~170K | **$0** |
-
-
-
-
-
-> Embedding model: `all-MiniLM-L6-v2` (lokal, libre, walang API key). Kumpletong mga ulat: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Paghambing sa kompetisyon: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) na sumasaklaw sa agentmemory laban sa mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**I-reproduce nang lokal:** [`eval/README.md`](../eval/README.md), isang adapter-pluggable harness para sa LongMemEval `_s` (public 500-Q) + `coding-agent-life-v1` (in-house na 15-session corpus). Nag-iiskor ang mga grep / vector / agentmemory adapter nang tabi-tabi, NDJSON output, dumadapo ang mga published scorecard sa [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Umaangkop sa [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything), at [Graphify](https://github.com/safishamsi/graphify).** Code-graph indexing, multi-agent build pipeline, at mas malawak na knowledge graph sa mga docs / PDF / imahe / video. Naaalala ng agentmemory ang ginawang trabaho; pinapaliwanag naman ng tatlong project na ito ang iba pang bahagi ng context layer. Mga recipe + question-routing table: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Built-in (CLAUDE.md)
-
-
-
Uri
-
Memory engine + MCP server
-
Memory layer API
-
Buong agent runtime
-
Personal AI
-
Memory API + app
-
Team memory hub (LLM proxy)
-
Vector memory (OSS)
-
Memory engine (Oracle DB)
-
Sistema ng memory
-
Static file
-
-
-
Retrieval R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Sariling ulat
-
PersonaMem 76% (sariling ulat)
-
~96.6% (sariling ulat)
-
94.4% (sariling ulat)
-
N/A
-
N/A (grep)
-
-
-
Auto-capture
-
12 hook (walang manu-manong pagsisikap)
-
Manual na add() call
-
Sariling pag-edit ng agent
-
Manual
-
API-side extraction
-
Proxy interception (pagpalit ng base-URL)
-
Manual
-
API extraction
-
Manual
-
Manual na pag-edit
-
-
-
Paghahanap
-
BM25 + Vector + Graph (RRF fusion)
-
Vector + Graph
-
Vector (archival)
-
Semantic
-
Vector + RAG
-
4 uri ng asset (Chat / Skill / Wiki / CodeGraph)
-
Vector lamang
-
Vector + semantic
-
Decay-weighted
-
Nilo-load ang lahat sa context
-
-
-
Multi-agent
-
MCP + REST + lease + signal
-
API (walang coordination)
-
Sa loob lamang ng Letta runtime
-
Hindi
-
Hindi
-
Mga tungkulin sa team + shared asset
-
Hindi
-
Scoped lamang
-
Shared sa multi-agent
-
Per-agent na file
-
-
-
Framework lock-in
-
Wala (kahit anong MCP client)
-
Wala
-
Mataas (kailangang gamitin ang Letta)
-
Standalone
-
Wala
-
Proxy sa harap ng bawat model call
-
Wala
-
Oracle Database
-
Wala
-
Per-agent na format
-
-
-
Panlabas na deps
-
Hindi kailangan ng hiwalay na database service
-
Qdrant / pgvector
-
Postgres + vector DB
-
Marami
-
Managed cloud
-
Docker stack (Core + Hub + Proxy)
-
Vector store
-
Oracle AI Database
-
Wala
-
Wala
-
-
-
Memory lifecycle
-
4-tier consolidation + decay + auto-forget
-
Passive extraction
-
Pinamamahalaan ng agent
-
Manual
-
Auto-forget
-
Manual review; isinasagawa pa ang auto-routing
-
Wala
-
Hindi nakasaad
-
Decay + consolidation
-
Manual na pruning
-
-
-
Kahusayan ng Token
-
~1,900 tokens/session ($10/yr)
-
Depende sa integration
-
Core memory sa context
-
Depende
-
Cloud pricing
-
Hindi nakasaad
-
Walang token budget
-
LLM-backed (depende)
-
Depende
-
22K+ tokens sa 240 obs
-
-
-
Real-time viewer
-
Oo (port 3113)
-
Cloud dashboard
-
Cloud dashboard
-
Web UI
-
Cloud dashboard
-
Hub web UI
-
Hindi
-
Hindi
-
Hindi
-
Hindi
-
-
-
Self-hosted
-
Oo (default)
-
Opsyonal
-
Opsyonal
-
Oo
-
Hindi (cloud-only)
-
Oo (Docker)
-
Oo
-
Oo (Oracle DB)
-
Oo
-
Oo
-
-
-
-Tala sa benchmark: ang R@5 lamang ng agentmemory ang aming sariling nasukat na resulta (LongMemEval-S, reproducible mula sa benchmark/COMPARISON.md). Ang mga figure ng mem0 at Letta ay ang kanilang mga pinublikang numero sa LoCoMo (ibang dataset); ang mga figure ng MemPalace, supermemory, TencentDB (PersonaMem), at oracleagentmemory ay mga claim na self-reported ng vendor na hindi namin independent na-reproduce (ang run ng oracleagentmemory ay gumamit ng GPT-5.5 laban sa isang Oracle AI Database). Ipinapakita nang tabi-tabi para sa ballpark estimate lamang, hindi isang head-to-head sa parehong data. Ang bilang ng star ay tinatantya at nagbabago sa paglipas ng panahon.
-
-**Mga bagong entrant** na dapat malaman, inihambing nang detalyado sa [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Sistema | ⭐ | Anggulo |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporal knowledge graph; ang pinakamalakas na pinublikang resulta sa temporal-query (LongMemEval 63.8%), pero bumubuo ang graph nang asynchronous kaya maaaring mahuli ang mga bagong fact |
-| Cognee | 30K | Document-to-knowledge-graph ingestion, Python lamang, ginawa para sa structured entity extraction sa halip na session capture |
+Na-update na paghahambing ng mga produkto at paraan ng deployment, kasama ang mga sanggunian:
-Pinagsasama ng AgentMemory ang session capture, searchable na project history, at local viewer, na magagamit nang walang key mula sa unang pagpapatakbo.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -780,7 +572,7 @@ Kumpletong gabay: [`integrations/hermes/`](../integrations/hermes/)
Simulan ang memory server: `npx -y @agentmemory/agentmemory@latest`
-#### Native na skill gamit ang `npx skills add` (50+ agent)
+#### Native na skill gamit ang `npx skills add`
Nag-ship ang agentmemory ng 17 skill sa Claude-Code-style na `/SKILL.md` format: 9 invocable na action skill (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) at 8 reference skill na nilo-load ng agent kapag kailangan (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). May dala ang mga reference skill na data table na nabuo mula sa source, kaya hindi ito kailanman nadrift. Ang [`skills`](https://npmjs.com/package/skills) CLI ng vercel-labs ay awtomatikong nag-install nito sa native skill directory ng tumatawag na agent sa 50+ agent (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, at iba pa):
@@ -837,7 +629,7 @@ Ang entry ng agentmemory ay ang **parehong MCP server block** sa bawat host na g
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | Ini-install ng `agentmemory connect antigravity --with-hooks` ang MCP at capture hooks sa shared customization directory. Tingnan ang [setup at limitasyon ng Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | Ginagamit ng `agentmemory connect antigravity-cli --with-hooks` ang parehong MCP at hook configuration ng kasalukuyang mga bersyon ng IDE. Dapat i-refresh ng mga umiiral na installation gamit ang `--force`; tingnan ang [upgrade notes](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | Isinusulat ng `agentmemory connect kiro` ang user-level config. Ang workspace override ay sa `.kiro/settings/mcp.json`, katabi ng iyong code. |
-| **Warp** | `~/.warp/.mcp.json` | Isinusulat ng `agentmemory connect warp` ang standard na `mcpServers` block. Awtomatikong nadidiskubre din ng Warp ang mga skill mula sa `.claude/skills/`; kapag naka-install na ang Claude Code plugin, lumalabas nang native ang 8 agentmemory skill (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) sa slash-command palette ng Warp. |
+| **Warp** | `~/.warp/.mcp.json` | Isinusulat ng `agentmemory connect warp` ang standard na `mcpServers` block. Awtomatikong nadidiskubre din ng Warp ang mga skill mula sa `.claude/skills/`; kapag naka-install na ang Claude Code plugin, lumalabas nang native ang 9 agentmemory skill (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) sa slash-command palette ng Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | Isinusulat ng `agentmemory connect cline` ang standard na `mcpServers` block. Para sa VS Code extension user: i-paste ang parehong block sa Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (preferred) o `config.json` (legacy) | Gumagawa ang `agentmemory connect continue` ng `config.yaml` kung wala pang configuration o ina-update ang kasalukuyang `config.json`. Hindi nito binabago ang umiiral na YAML upang mapanatili ang comments at anchors; ipinapakita ang block na ilalagay sa ilalim ng `mcpServers:`. Array ang gamit ng Continue para sa `mcpServers`. |
| **Zed** | `~/.config/zed/settings.json` | Isinusulat ng `agentmemory connect zed` sa ilalim ng `context_servers` (ang key ng Zed, HINDI `mcpServers`). Maaaring i-wire ang remote MCP server gamit ang `{"url": "..."}` sa halip. |
@@ -933,7 +725,6 @@ Nag-ship ang bawat AI coding agent na may built-in memory: may `MEMORY.md` ang C
|---|---|---|
| Pagpapanatili | Manu-manong inayos na instructions | Naka-index na session history at memory |
| Paghahanap | Nilo-load ang lahat sa context | BM25 + vector + graph (top-K lamang) |
-| Gastos sa Token | 22K+ sa 240 obserbasyon | ~1,900 token (92% mas kaunti) |
| Cross-agent | Per-agent na file | MCP + REST (kahit anong agent) |
| Koordinasyon | Wala | Lease, signal, action, routine |
| Observability | Basahin ang mga file nang manu-mano | Real-time viewer sa :3113 |
@@ -946,7 +737,7 @@ Nag-ship ang bawat AI coding agent na may built-in memory: may `MEMORY.md` ang C
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1031,7 +822,7 @@ Pinagsasama gamit ang Reciprocal Rank Fusion (RRF, k=60) at session-diversified
Kapag napupuno ang isang vector index, ginagamit ng `mem::search` (sa likod ng `memory_recall`) ang hybrid na BM25 + vector ranker. Kapag walang embeddings, ginagamit nito ang BM25. Maaari pang pagsamahin ng `smart-search` ang structural graph matches kapag umiiral ang graph data, kasama na sa keyless mode. Tumatakbo ang lesson recall sa isang dedikadong in-memory BM25 index sa halip na i-scan ang buong corpus kada query. Hindi kasama sa bawat recall path ang mga superseded na bersyon ng memory; pinapanatili ng version chain ang kanilang history.
-Nakaligtas ang mga vector sa isang crash o force-kill. Nasa-save ang vector index sa mga bucket sa pinakamadalas na bawat `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 minuto). Bawat vector na idinagdag o inalis sa gitna ay agad na isinusulat din sa isang maliit na pending log sa state store, at ni-replay ito ng susunod na start nang hindi tinawag ang embedding provider. Ang bawat matagumpay na save ay nagwawalang-laman sa log. Ang mga documento na wala pa ring vector pagkatapos ng replay ay re-embedded sa background sa mga batch ng `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) hanggang maubos ang lahat, at ang backfill na natigil ay magpapatuloy sa susunod na start. Ipinapakita ng `/agentmemory/status` at ng viewer ang laki ng pending log at ang estado ng backfill. Walang isinusulat ang mga keyless install.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
Tinotokenize ng BM25 ang Griyego, Cyrillic, Hebreo, Arabic, at accented Latin nang out of the box. Para sa mga memory sa Chinese / Japanese / Korean, i-install ang opsyonal na segmenter (`npm install @node-rs/jieba tiny-segmenter`) para hatiin ang mga CJK run sa word-level token; kung wala ito, soft-falls ang agentmemory sa whole-run tokenization at nagprint ng one-time hint sa stderr.
@@ -1047,9 +838,9 @@ Kasama sa normal na npm install ang opsyonal na `@huggingface/transformers` runt
| Provider | Model | Gastos | Mga Tala |
|---|---|---|---|
-| **Local (inirerekomendang opt-in)** | `all-MiniLM-L6-v2` | Libre | On-device pagkatapos ng unang pag-download ng model, +8pp recall kaysa sa BM25-only |
+| **Local (inirerekomendang opt-in)** | `all-MiniLM-L6-v2` | Libre | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Free tier | 100+ wika, 768/1536/3072 dims (MRL), 2048-token input. Pumalit sa `text-embedding-004` ([deprecated, magsasara noong Jan 14, 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Pinakamataas na kalidad |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Bayad | Na-optimize para sa code |
| Cohere | `embed-english-v3.0` | Free trial | Pangkalahatang gamit |
| OpenRouter | Kahit anong model | Depende | Multi-model proxy |
diff --git a/READMEs/README.tr-TR.md b/READMEs/README.tr-TR.md
index 014dd3f7ea..0f8aa10db7 100644
--- a/READMEs/README.tr-TR.md
+++ b/READMEs/README.tr-TR.md
@@ -66,8 +66,6 @@
-
-
@@ -333,223 +331,17 @@ En son sürüm notları: [CHANGELOG.md](../CHANGELOG.md).
-
-
-
-
-### Geri Getirme Doğruluğu
-
-**coding-agent-life-v1** (şirket içi (in-house) corpus, sandbox'ta yeniden üretilebilir)
-
-| Adaptör | P@5 | R@5 | İlk 5'te bulma oranı | p50 gecikme |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep baseline | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-Bu corpus için **P@5 matematiksel tavanında** (0.240, scorecard'a bakın) %100 ilk-5 bulma oranı. Hybrid, her gold oturumu getirir; grep, çok-oturumlu temporal sorguda 2 gold'dan 1'ini kaçırır. Kazanım **recall + temporal**'dir, toplam precision değil. Bu benchmark küçüktür ve gold açısından seyrektir; aşağıdaki daha büyük LongMemEval-S daha iyi ayrıştırır. Tam tür bazlı döküm + düzeltme notu: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 soru)
+Güncel bilgi getirme sonuçları, metrik tanımları ve yeniden üretme adımları:
-| Sistem | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| Yalnızca BM25 fallback | 86.2% | 94.6% | 71.5% |
-
-
-
-> Embedding modeli: `all-MiniLM-L6-v2` (yerel, ücretsiz, API anahtarı gerektirmez). Tam raporlar: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Rakip karşılaştırması: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) — agentmemory'yi mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo ile karşılaştırır.
-
-**Yerelde yeniden üretin:** [`eval/README.md`](../eval/README.md), LongMemEval `_s` (herkese açık 500 soru) + `coding-agent-life-v1` (şirket içi 15-oturumluk corpus) için adaptör-takılabilir bir harness. Grep / vektör / agentmemory adaptörleri yan yana puanlanır, NDJSON çıktısı verir, yayımlanan scorecard'lar [`docs/benchmarks/`](../docs/benchmarks/) içine düşer.
-
-**[codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) ve [Graphify](https://github.com/safishamsi/graphify) ile eşleşir.** Kod-grafı indeksleme, çok-ajanlı build pipeline'ları ve doküman / PDF / görüntü / video genelinde daha geniş bilgi grafları. agentmemory yapılan işi hatırlar; bu üç proje, bağlam katmanının gerisini aydınlatır. Tarifler + soru-yönlendirme tablosu: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Yerleşik (CLAUDE.md)
-
-
-
Tür
-
Bellek engine'i + MCP sunucusu
-
Bellek katmanı API'si
-
Tam ajan runtime'ı
-
Kişisel AI
-
Bellek API'si + uygulama
-
Takım belleği hub'ı (LLM proxy)
-
Vektör belleği (OSS)
-
Bellek engine'i (Oracle DB)
-
Bellek sistemi
-
Statik dosya
-
-
-
Geri Getirme R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Kendi beyanı
-
PersonaMem 76% (kendi beyanı)
-
~96.6% (kendi beyanı)
-
94.4% (kendi beyanı)
-
N/A
-
N/A (grep)
-
-
-
Otomatik yakalama
-
12 hook (sıfır manuel çaba)
-
Manuel add() çağrıları
-
Ajan kendi kendine düzenler
-
Manuel
-
API tarafında çıkarım
-
Proxy ile yakalama (base-URL değişimi)
-
Manuel
-
API çıkarımı
-
Manuel
-
Manuel düzenleme
-
-
-
Arama
-
BM25 + Vektör + Graf (RRF füzyonu)
-
Vektör + Graf
-
Vektör (arşivsel)
-
Semantik
-
Vektör + RAG
-
4 varlık türü (Chat / Skill / Wiki / CodeGraph)
-
Yalnızca vektör
-
Vektör + semantik
-
Çürüme ağırlıklı (decay-weighted)
-
Her şeyi bağlama yükler
-
-
-
Çoklu ajan
-
MCP + REST + lease'ler + signal'ler
-
API (koordinasyon yok)
-
Yalnızca Letta runtime'ı içinde
-
Hayır
-
Hayır
-
Takım rolleri + paylaşılan varlıklar
-
Hayır
-
Yalnızca scope'lu
-
Çoklu ajan arasında paylaşılan
-
Ajan başına dosyalar
-
-
-
Framework bağımlılığı (lock-in)
-
Yok (herhangi bir MCP istemcisi)
-
Yok
-
Yüksek (Letta kullanmak zorunlu)
-
Bağımsız
-
Yok
-
Proxy her model çağrısının önünde
-
Yok
-
Oracle Database
-
Yok
-
Ajan başına format
-
-
-
Harici bağımlılıklar
-
Ayrı bir veritabanı hizmeti gerektirmez
-
Qdrant / pgvector
-
Postgres + vektör DB
-
Birden fazla
-
Yönetilen bulut
-
Docker stack'i (Core + Hub + Proxy)
-
Vektör deposu
-
Oracle AI Database
-
Yok
-
Yok
-
-
-
Bellek yaşam döngüsü
-
4 katmanlı konsolidasyon + çürüme + otomatik unutma
-
Pasif çıkarım
-
Ajan tarafından yönetilen
-
Manuel
-
Otomatik unutma
-
Manuel inceleme; otomatik yönlendirme geliştiriliyor
-
Yok
-
Belirtilmemiş
-
Çürüme + konsolidasyon
-
Manuel budama
-
-
-
Token verimliliği
-
~1,900 token/oturum ($10/yıl)
-
Entegrasyona göre değişir
-
Çekirdek bellek bağlamda
-
Değişir
-
Bulut fiyatlandırması
-
Belirtilmemiş
-
Token bütçesi yok
-
LLM destekli (değişir)
-
Değişir
-
240 gözlemde 22K+ token
-
-
-
Gerçek zamanlı görüntüleyici
-
Evet (port 3113)
-
Bulut dashboard'u
-
Bulut dashboard'u
-
Web arayüzü
-
Bulut dashboard'u
-
Hub web arayüzü
-
Hayır
-
Hayır
-
Hayır
-
Hayır
-
-
-
Kendi sunucunda barındırma
-
Evet (varsayılan)
-
Opsiyonel
-
Opsiyonel
-
Evet
-
Hayır (yalnızca bulut)
-
Evet (Docker)
-
Evet
-
Evet (Oracle DB)
-
Evet
-
Evet
-
-
-
-Benchmark notu: yalnızca agentmemory'nin R@5'i bizim kendi ölçtüğümüz sonuçtur (LongMemEval-S, benchmark/COMPARISON.md'dan yeniden üretilebilir). mem0 ve Letta rakamları, onların yayımladığı LoCoMo sayılarıdır (farklı bir veri kümesi); MemPalace, supermemory, TencentDB (PersonaMem) ve oracleagentmemory rakamları, bağımsız olarak yeniden üretmediğimiz, satıcı tarafından kendi beyan edilen iddialardır (oracleagentmemory'nin çalıştırması bir Oracle AI Database'e karşı GPT-5.5 kullandı). Yalnızca kabaca karşılaştırma için yan yana gösterilmiştir, aynı veri üzerinde doğrudan bir karşılaştırma değildir. Yıldız sayıları yaklaşıktır ve zamanla değişir.
-
-**Bilinmesi gereken daha yeni oyuncular**, [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) içinde derinlemesine karşılaştırılmıştır:
-
-| Sistem | ⭐ | Yaklaşım |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporal bilgi grafı; yayımlanmış en güçlü temporal-sorgu sonuçları (LongMemEval 63.8%), ancak graf asenkron olarak oluşturulur, bu yüzden yeni bilgiler gecikebilir |
-| Cognee | 30K | Doküman-to-bilgi-grafı alımı, yalnızca Python, oturum yakalamadan ziyade yapılandırılmış entity çıkarımı için inşa edilmiştir |
+Kaynaklarıyla birlikte ürünlerin ve dağıtım seçeneklerinin güncel karşılaştırması:
-AgentMemory oturum kaydını, aranabilir proje geçmişini ve yerel görüntüleyiciyi birleştirir; anahtarsız çalışma ilk başlatmadan itibaren kullanılabilir.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -786,7 +578,7 @@ Tam kılavuz: [`integrations/hermes/`](../integrations/hermes/)
Bellek sunucusunu başlatın: `npx -y @agentmemory/agentmemory@latest`
-#### `npx skills add` ile yerel skill'ler (50+ ajan)
+#### `npx skills add` ile yerel skill'ler
agentmemory, Claude-Code-tarzı `/SKILL.md` formatında 17 skill sunar: 9 çağrılabilir aksiyon skill'i (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) ve ajanın gerektiğinde yüklediği 8 referans skill (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Referans skill'ler, kaynaktan üretilen veri tablolarını taşır, bu yüzden asla sapmazlar. vercel-labs'ın [`skills`](https://npmjs.com/package/skills) CLI'si, bunları çağıran ajanın yerel skill dizinine 50+ ajanda (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf ve daha fazlası) otomatik olarak kurar:
@@ -844,7 +636,7 @@ agentmemory girdisi, `mcpServers` şeklini kullanan her host'ta (Cursor, Claude
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks`, paylaşılan customization dizinine MCP ve yakalama hook'larını kurar. Bkz. [Antigravity kurulumu ve sınırları](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks`, güncel IDE sürümleriyle aynı MCP ve hook yapılandırmasını kullanır. Mevcut kurulumlar `--force` ile yenilenmelidir; bkz. [yükseltme notları](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro`, kullanıcı seviyesindeki yapılandırmayı yazar. Workspace geçersiz kılmaları, kodunuzun yanındaki `.kiro/settings/mcp.json` içine gider. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp`, standart `mcpServers` bloğunu yazar. Warp, ayrıca `.claude/skills/`'ten skill'leri otomatik keşfeder; Claude Code eklentisi kurulduğunda 8 agentmemory skill'i (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`), Warp'ın slash-komut paletinde yerli olarak görünür. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp`, standart `mcpServers` bloğunu yazar. Warp, ayrıca `.claude/skills/`'ten skill'leri otomatik keşfeder; Claude Code eklentisi kurulduğunda 9 agentmemory skill'i (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`), Warp'ın slash-komut paletinde yerli olarak görünür. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline`, standart `mcpServers` bloğunu yazar. VS Code eklenti kullanıcıları: aynı bloğu Cline Settings → MCP Servers → Edit JSON üzerinden yapıştırın. |
| **Continue.dev** | `~/.continue/config.yaml` (tercih edilen) veya `config.json` (eski/legacy) | `agentmemory connect continue`, yapılandırma yoksa `config.yaml` oluşturur veya mevcut `config.json` dosyasını değiştirir. Mevcut YAML için yorumları ve çapaları koruyarak `mcpServers:` altına eklenecek bloğu gösterir. Continue nesne yerine dizi kullanır. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed`, `context_servers` altına yazar (Zed'in key'i, `mcpServers` DEĞİL). Uzak MCP sunucuları, bunun yerine `{"url": "..."}` üzerinden bağlanabilir. |
@@ -926,7 +718,6 @@ Her AI kodlama ajanı, yerleşik bellekle gelir: Claude Code'da `MEMORY.md`, Cur
|---|---|---|
| Bakım | Elle düzenlenen talimatlar | İndekslenmiş oturum geçmişi ve anılar |
| Arama | Her şeyi bağlama yükler | BM25 + vektör + graf (yalnızca top-K) |
-| Token maliyeti | 240 gözlemde 22K+ | ~1,900 token (92% daha az) |
| Ajanlar arası | Ajan başına dosyalar | MCP + REST (herhangi bir ajan) |
| Koordinasyon | Yok | Lease'ler, signal'ler, action'lar, routine'ler |
| Gözlemlenebilirlik | Dosyaları manuel okuma | :3113'te gerçek zamanlı görüntüleyici |
@@ -939,7 +730,7 @@ Her AI kodlama ajanı, yerleşik bellekle gelir: Claude Code'da `MEMORY.md`, Cur
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1024,7 +815,7 @@ Reciprocal Rank Fusion (RRF, k=60) ile füzyonlanır ve oturum bazında çeşitl
Bir vektör indeksi doldurulduğunda, `mem::search` (`memory_recall`'un arkasında) hybrid BM25 + vektör ranker'ı kullanır. Embedding olmadan BM25 kullanır. `smart-search`, anahtarsız modda dahi graf verisi mevcut olduğunda ek olarak yapısal graf eşleşmelerini füzyonlayabilir. Lesson recall'u, her sorguda tüm corpus'u taramak yerine özel bir bellek içi (in-memory) BM25 indeksinde çalışır. Supersede edilmiş bellek sürümleri her recall yolundan hariç tutulur; sürüm zinciri geçmişlerini korur.
-Vektörler bir çökme veya force-kill'den sağ çıkar. Vektör indeksi, en fazla her `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS`'de (10 dakika) bucket'lar halinde kaydedilir. Bu aralıkta eklenen veya kaldırılan her vektör, state store'daki küçük bir bekleyen (pending) log'a da hemen yazılır ve bir sonraki başlatma, embedding sağlayıcısını çağırmadan bunu replay eder. Her başarılı kayıt, log'u boşaltır. Replay'den sonra hâlâ vektörü olmayan dokümanlar, hiçbiri kalmayana kadar `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) parti halinde arka planda yeniden embed edilir ve durdurulan bir backfill, bir sonraki başlatmada devam eder. `/agentmemory/status` ve görüntüleyici, bekleyen log boyutunu ve backfill durumunu gösterir. Anahtarsız kurulumlar hiçbir şey yazmaz.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25, Yunanca, Kiril, İbranice, Arapça ve aksanlı Latin'i kutudan çıktığı gibi tokenize eder. Çince / Japonca / Korece bellekler için, CJK run'larını kelime seviyesinde token'lara bölmek üzere opsiyonel segmenter'ları kurun (`npm install @node-rs/jieba tiny-segmenter`); bunlar olmadan agentmemory, tüm-run tokenizasyonuna yumuşak geçiş yapar (soft-falls) ve stderr'de tek seferlik bir ipucu yazdırır.
@@ -1040,9 +831,9 @@ Normal npm kurulumu, opsiyonel `@huggingface/transformers` runtime'ını içerir
| Sağlayıcı | Model | Maliyet | Notlar |
|---|---|---|---|
-| **Yerel (önerilen opt-in)** | `all-MiniLM-L6-v2` | Ücretsiz | İlk model indirmesinden sonra cihaz üzerinde, yalnızca BM25'e göre +8pp recall |
+| **Yerel (önerilen opt-in)** | `all-MiniLM-L6-v2` | Ücretsiz | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Ücretsiz katman | 100+ dil, 768/1536/3072 boyut (MRL), 2048-token girdi. `text-embedding-004`'ün yerini alır ([kullanımdan kaldırıldı, 14 Oca 2026'da durduruldu](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | En yüksek kalite |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Ücretli | Kod için optimize edilmiş |
| Cohere | `embed-english-v3.0` | Ücretsiz deneme | Genel amaçlı |
| OpenRouter | Herhangi bir model | Değişken | Çoklu-model proxy |
diff --git a/READMEs/README.uk-UA.md b/READMEs/README.uk-UA.md
index 1f142d4858..2c67111fe3 100644
--- a/READMEs/README.uk-UA.md
+++ b/READMEs/README.uk-UA.md
@@ -66,8 +66,6 @@
-
-### Точність пригадування
-
-**coding-agent-life-v1** (власний корпус, відтворюваний у sandbox)
-
-| Адаптер | P@5 | R@5 | Частка влучань у топ-5 | латентність p50 |
-|---|---|---|---|---|
-| **agentmemory (гібридний)** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| базовий варіант grep | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-100% влучань у топ-5 на рівні **математичної межі P@5** для цього корпусу (0.240, див. scorecard). Гібридний режим знаходить усі золоті сесії; grep пропускає 1 з 2 золотих результатів у мультисесійному темпоральному запиті. Прирощення — це **пригадування + темпоральність**, а не агрегована точність. Цей бенчмарк невеликий і має мало золотих прикладів; більший LongMemEval-S нижче краще розрізняє результати. Повна розбивка за типами + примітка про виправлення: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 запитань)
+Актуальні результати пошуку, визначення метрик і кроки відтворення:
-| Система | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| резервний варіант лише на BM25 | 86.2% | 94.6% | 71.5% |
-
-
-
-> Модель ембедингів: `all-MiniLM-L6-v2` (локальна, безкоштовна, без API-ключа). Повні звіти: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). Порівняння з конкурентами: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) охоплює agentmemory порівняно з mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Відтворити локально:** [`eval/README.md`](../eval/README.md) — придатний до підключення адаптерів harness для LongMemEval `_s` (публічний набір із 500 запитань) + `coding-agent-life-v1` (власний корпус із 15 сесій). Адаптери grep / vector / agentmemory оцінюються поруч, вивід у форматі NDJSON, опубліковані scorecard потрапляють у [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Добре поєднується з [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything) та [Graphify](https://github.com/safishamsi/graphify).** Індексування графа коду, конвеєри збирання для кількох агентів і ширші графи знань за документами / PDF / зображеннями / відео. agentmemory запам'ятовує роботу; ці три проєкти оживляють решту контекстного шару. Рецепти + таблиця маршрутизації питань: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Вбудована (CLAUDE.md)
-
-
-
Тип
-
Двигун пам'яті + MCP-сервер
-
API шару пам'яті
-
Повноцінний рантайм агента
-
Персональний AI
-
API пам'яті + застосунок
-
Хаб командної пам'яті (LLM-проксі)
-
Векторна пам'ять (OSS)
-
Двигун пам'яті (Oracle DB)
-
Система пам'яті
-
Статичний файл
-
-
-
Пригадування R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
Н/Д
-
Заявлено самостійно
-
PersonaMem 76% (заявлено самостійно)
-
~96.6% (заявлено самостійно)
-
94.4% (заявлено самостійно)
-
Н/Д
-
Н/Д (grep)
-
-
-
Автозахоплення
-
12 хуків (без ручних дій)
-
Ручні виклики add()
-
Самостійне редагування агентом
-
Вручну
-
Видобування на стороні API
-
Перехоплення через проксі (підміна base-URL)
-
Вручну
-
Видобування через API
-
Вручну
-
Ручне редагування
-
-
-
Пошук
-
BM25 + вектор + граф (злиття RRF)
-
Вектор + граф
-
Вектор (архівний)
-
Семантичний
-
Вектор + RAG
-
4 типи активів (Chat / Skill / Wiki / CodeGraph)
-
Лише вектор
-
Вектор + семантичний
-
Зважений за згасанням
-
Завантажує все у контекст
-
-
-
Мультиагентність
-
MCP + REST + оренди + сигнали
-
API (без координації)
-
Лише в межах рантайму Letta
-
Ні
-
Ні
-
Командні ролі + спільні активи
-
Ні
-
Лише в межах скоупу
-
Спільна для кількох агентів
-
Файли для кожного агента
-
-
-
Прив'язка до фреймворку
-
Немає (будь-який MCP-клієнт)
-
Немає
-
Висока (потрібно використовувати Letta)
-
Автономний
-
Немає
-
Проксі обробляє кожен виклик моделі
-
Немає
-
Oracle Database
-
Немає
-
Формат для кожного агента
-
-
-
Зовнішні залежності
-
Окрема служба бази даних не потрібна
-
Qdrant / pgvector
-
Postgres + векторна БД
-
Декілька
-
Керована хмара
-
Docker-стек (Core + Hub + Proxy)
-
Векторне сховище
-
Oracle AI Database
-
Немає
-
Немає
-
-
-
Життєвий цикл пам'яті
-
4-рівнева консолідація + згасання + автозабування
-
Пасивне видобування
-
Керується агентом
-
Вручну
-
Автозабування
-
Ручний перегляд; автоматична маршрутизація в розробці
-
Немає
-
Не вказано
-
Згасання + консолідація
-
Ручне очищення
-
-
-
Ефективність токенів
-
~1,900 токенів/сесію ($10/рік)
-
Залежить від інтеграції
-
Основна пам'ять у контексті
-
Залежить
-
Хмарне ціноутворення
-
Не вказано
-
Без бюджету токенів
-
На основі LLM (залежить)
-
Залежить
-
22K+ токенів при 240 спостереженнях
-
-
-
Переглядач у реальному часі
-
Так (порт 3113)
-
Хмарна панель
-
Хмарна панель
-
Веб-інтерфейс
-
Хмарна панель
-
Веб-інтерфейс Hub
-
Ні
-
Ні
-
Ні
-
Ні
-
-
-
Самостійний хостинг
-
Так (за замовчуванням)
-
Опційно
-
Опційно
-
Так
-
Ні (лише хмара)
-
Так (Docker)
-
Так
-
Так (Oracle DB)
-
Так
-
Так
-
-
-
-Примітка до бенчмарків: лише R@5 для agentmemory — це наш власний вимірений результат (LongMemEval-S, відтворюваний з benchmark/COMPARISON.md). Показники mem0 і Letta — це їхні опубліковані числа LoCoMo (інший набір даних); показники MemPalace, supermemory, TencentDB (PersonaMem) та oracleagentmemory — це заявлені самими постачальниками твердження, які ми не відтворювали самостійно (запуск oracleagentmemory використовував GPT-5.5 проти Oracle AI Database). Наведено поруч лише для орієнтовного порівняння, а не як пряме протистояння на однакових даних. Кількість зірок приблизна і змінюється з часом.
-
-**Новіші учасники**, варті уваги, детально порівняні в [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Система | ⭐ | Підхід |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Темпоральний граф знань; найкращі опубліковані результати для темпоральних запитів (LongMemEval 63.8%), але граф будується асинхронно, тому нові факти можуть надходити із затримкою |
-| Cognee | 30K | Перетворення документів на граф знань, лише Python, створено для структурованого видобування сутностей, а не для захоплення сесій |
+Оновлене порівняння продуктів і варіантів розгортання з джерелами:
-AgentMemory поєднує захоплення сесій, пошук в історії проєкту й локальний viewer; робота без ключа доступна з першого запуску.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -786,7 +578,7 @@ Verify with `curl http://localhost:3111/agentmemory/health`. Open http://localho
Запустіть сервер пам'яті: `npx -y @agentmemory/agentmemory@latest`
-#### Нативні скіли через `npx skills add` (50+ агентів)
+#### Нативні скіли через `npx skills add`
agentmemory постачає 17 скілів у форматі `/SKILL.md` у стилі Claude Code: 9 викликаних скілів-дій (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) і 8 довідкових скілів, які агент завантажує за потреби (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Довідкові скіли містять таблиці даних, згенеровані з джерела, тому вони ніколи не розходяться з ним. CLI [`skills`](https://npmjs.com/package/skills) від vercel-labs автоматично встановлює їх у нативний каталог скілів агента, що викликає, для понад 50 агентів (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf та інших):
@@ -844,7 +636,7 @@ npx skills add rohitg00/agentmemory -y -a '*' # install to every installed age
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` встановлює MCP і хуки захоплення у спільному каталозі налаштувань. Див. [налаштування та обмеження Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` використовує ту саму конфігурацію MCP і хуків, що й поточні версії IDE. Наявні встановлення слід оновити за допомогою `--force`; див. [примітки щодо оновлення](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` записує конфігурацію рівня користувача. Перевизначення на рівні робочого простору розміщуються в `.kiro/settings/mcp.json` поруч із вашим кодом. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` записує стандартний блок `mcpServers`. Warp також автоматично виявляє скіли з `.claude/skills/`; після встановлення плагіна Claude Code 8 скілів agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) з'являються нативно в палітрі слеш-команд Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` записує стандартний блок `mcpServers`. Warp також автоматично виявляє скіли з `.claude/skills/`; після встановлення плагіна Claude Code 9 скілів agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) з'являються нативно в палітрі слеш-команд Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` записує стандартний блок `mcpServers`. Користувачі розширення VS Code: вставте той самий блок через Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (бажано) або `config.json` (застарілий) | `agentmemory connect continue` створює `config.yaml`, якщо налаштувань немає, або змінює наявний `config.json`. Для наявного YAML показує блок для `mcpServers:`, зберігаючи коментарі та якорі. Continue використовує масив, а не об’єкт. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` записує дані під `context_servers` (ключ Zed, НЕ `mcpServers`). Віддалені MCP-сервери замість цього можна підключити через `{"url": "..."}`. |
@@ -927,7 +719,6 @@ REST використовує `3111`, viewer `3113`. Залиште їх у пр
|---|---|---|
| Обслуговування | Інструкції, редаговані вручну | Індексована історія сесій і спогади |
| Пошук | Завантажує все у контекст | BM25 + вектор + граф (лише топ-K) |
-| Вартість у токенах | 22K+ при 240 спостереженнях | ~1,900 токенів (на 92% менше) |
| Між агентами | Файли для кожного агента | MCP + REST (будь-який агент) |
| Координація | Немає | Оренди, сигнали, дії, рутини |
| Спостережуваність | Читання файлів вручну | Переглядач у реальному часі на :3113 |
@@ -940,7 +731,7 @@ REST використовує `3111`, viewer `3113`. Залиште їх у пр
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1025,7 +816,7 @@ SessionStart hook fires
Коли векторний індекс заповнено, `mem::search` (що стоїть за `memory_recall`) використовує гібридний ранжувальник BM25 + вектор. Без ембедингів він використовує BM25. `smart-search` може додатково об'єднувати структурні збіги графа, якщо дані графа існують, зокрема в режимі без ключа. Пригадування уроків виконується на окремому BM25-індексі в пам'яті, а не шляхом сканування всього корпусу під час кожного запиту. Заміщені версії пам'яті виключаються з усіх шляхів пригадування; ланцюжок версій зберігає їхню історію.
-Вектори переживають збій або примусове завершення. Векторний індекс зберігається пакетами не частіше, ніж кожні `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 хвилин). Кожен вектор, доданий або видалений у проміжку, також відразу записується у невеликий журнал очікування в хранилищі стану, і наступний запуск відтворює його без звернення до постачальника ембедингів. Кожне успішне збереження очищає журнал. Документи, які все ще не мають вектора після відтворення, повторно ембедяться у фоновому режимі партіями по `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500), доки не залишиться жодного, а зупинене доповнення продовжується з наступного запуску. `/agentmemory/status` і переглядач показують розмір журналу очікування та стан доповнення. Встановлення без ключа нічого не записують.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 одразу токенізує грецьку, кирилицю, іврит, арабську та латиницю з діакритикою. Для пам'яті китайською / японською / корейською встановіть опційні сегментатори (`npm install @node-rs/jieba tiny-segmenter`), щоб розбивати послідовності CJK на токени рівня слів; без них agentmemory плавно переходить на токенізацію цілими послідовностями та одноразово виводить підказку в stderr.
@@ -1041,9 +832,9 @@ EMBEDDING_PROVIDER=local
| Постачальник | Модель | Вартість | Примітки |
|---|---|---|---|
-| **Локальний (рекомендоване підключення)** | `all-MiniLM-L6-v2` | Безкоштовно | На пристрої після першого завантаження моделі, +8пп до пригадування порівняно з лише BM25 |
+| **Локальний (рекомендоване підключення)** | `all-MiniLM-L6-v2` | Безкоштовно | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Безкоштовний тариф | 100+ мов, 768/1536/3072 вимірів (MRL), вхід до 2048 токенів. Замінює `text-embedding-004` ([застаріла, вимкнення 14 січня 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Найвища якість |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Платно | Оптимізовано для коду |
| Cohere | `embed-english-v3.0` | Безкоштовна пробна версія | Загального призначення |
| OpenRouter | Будь-яка модель | Залежить | Проксі для декількох моделей |
diff --git a/READMEs/README.ur-PK.md b/READMEs/README.ur-PK.md
index 9ea6efafdc..f16fee5a18 100644
--- a/READMEs/README.ur-PK.md
+++ b/READMEs/README.ur-PK.md
@@ -68,8 +68,6 @@
-
-> Embedding ماڈل: `all-MiniLM-L6-v2` (مقامی، مفت، کوئی API key نہیں)۔ مکمل رپورٹس: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md)، [`benchmark/QUALITY.md`](../benchmark/QUALITY.md)، [`benchmark/SCALE.md`](../benchmark/SCALE.md)۔ حریف موازنہ: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) جو agentmemory کا mem0، Letta، Khoj، supermemory، TencentDB Agent Memory، MemPalace، Zep/Graphiti، Cognee، Hippo سے موازنہ کرتا ہے۔
-
-**مقامی طور پر دہرائیں:** [`eval/README.md`](../eval/README.md)، LongMemEval `_s` (عوامی 500-سوالات) + `coding-agent-life-v1` (اندرونی 15-سیشن corpus) کے لیے ایک adapter-pluggable ہارنس۔ Grep / vector / agentmemory adapters ساتھ ساتھ اسکور کرتے ہیں، NDJSON آؤٹ پٹ، شائع شدہ scorecards [`docs/benchmarks/`](../docs/benchmarks/) میں آتے ہیں۔
-
-**[codegraph](https://github.com/colbymchenry/codegraph)، [Understand Anything](https://github.com/Lum1104/Understand-Anything)، اور [Graphify](https://github.com/safishamsi/graphify) کے ساتھ جوڑی بناتا ہے۔** Code-graph انڈیکسنگ، multi-agent build pipelines، اور docs / PDFs / images / videos میں وسیع نالج گرافس۔ agentmemory کام کو یاد رکھتا ہے؛ یہ تین پراجیکٹس باقی context layer کو روشن کرتے ہیں۔ Recipes + سوال-روٹنگ ٹیبل: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md)۔
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
بلٹ-اِن (CLAUDE.md)
-
-
-
قسم
-
میموری انجن + MCP سرور
-
میموری لیئر API
-
مکمل ایجنٹ رن ٹائم
-
ذاتی AI
-
میموری API + ایپ
-
ٹیم میموری ہب (LLM پراکسی)
-
ویکٹر میموری (OSS)
-
میموری انجن (Oracle DB)
-
میموری سسٹم
-
سٹیٹک فائل
-
-
-
ریٹریول R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
خود رپورٹ شدہ
-
PersonaMem 76% (خود رپورٹ شدہ)
-
~96.6% (خود رپورٹ شدہ)
-
94.4% (خود رپورٹ شدہ)
-
N/A
-
N/A (grep)
-
-
-
خودکار کیپچر
-
12 ہکس (کوئی دستی کام نہیں)
-
دستی `add()` کالز
-
ایجنٹ خود ترمیم کرتا ہے
-
دستی
-
API کی جانب سے استخراج
-
پراکسی انٹرسیپشن (base-URL سویپ)
-
دستی
-
API استخراج
-
دستی
-
دستی ترمیم
-
-
-
سرچ
-
BM25 + ویکٹر + گراف (RRF فیوژن)
-
ویکٹر + گراف
-
ویکٹر (آرکائیول)
-
سیمینٹک
-
ویکٹر + RAG
-
4 asset اقسام (Chat / Skill / Wiki / CodeGraph)
-
صرف ویکٹر
-
ویکٹر + سیمینٹک
-
زوال پر مبنی وزن
-
سب کچھ context میں لوڈ کرتا ہے
-
-
-
ملٹی-ایجنٹ
-
MCP + REST + لیز + سگنلز
-
API (کوئی ربط نہیں)
-
صرف Letta رن ٹائم کے اندر
-
نہیں
-
نہیں
-
ٹیم کردار + مشترک asset
-
نہیں
-
صرف محدود دائرہ کار میں
-
ملٹی-ایجنٹ مشترک
-
فی-ایجنٹ فائلیں
-
-
-
فریم ورک لاک-اِن
-
کوئی نہیں (کوئی بھی MCP کلائنٹ)
-
کوئی نہیں
-
زیادہ (لازمی طور پر Letta استعمال کرنا ہوگا)
-
اسٹینڈ الون
-
کوئی نہیں
-
ہر model کال کے سامنے پراکسی ہوتی ہے
-
کوئی نہیں
-
Oracle Database
-
کوئی نہیں
-
فی-ایجنٹ فارمیٹ
-
-
-
خارجی انحصار
-
الگ ڈیٹابیس سروس درکار نہیں
-
Qdrant / pgvector
-
Postgres + ویکٹر DB
-
متعدد
-
مینیجڈ کلاؤڈ
-
Docker اسٹیک (Core + Hub + Proxy)
-
ویکٹر اسٹور
-
Oracle AI Database
-
کوئی نہیں
-
کوئی نہیں
-
-
-
میموری لائف سائیکل
-
4-درجاتی استحکام + زوال + خودکار فراموشی
-
غیر فعال استخراج
-
ایجنٹ کے زیرِ انتظام
-
دستی
-
خودکار فراموشی
-
دستی جائزہ؛ auto-routing زیرِ تکمیل
-
کوئی نہیں
-
واضح نہیں
-
زوال + استحکام
-
دستی تراش
-
-
-
ٹوکن کارکردگی
-
~1,900 ٹوکنز/سیشن ($10/سال)
-
integration کے لحاظ سے مختلف
-
بنیادی میموری context میں
-
مختلف ہوتا ہے
-
کلاؤڈ قیمتیں
-
واضح نہیں
-
کوئی token بجٹ نہیں
-
LLM پر مبنی (مختلف ہوتا ہے)
-
مختلف ہوتا ہے
-
240 obs پر 22K+ ٹوکنز
-
-
-
ریئل-ٹائم ویور
-
ہاں (پورٹ 3113)
-
کلاؤڈ ڈیش بورڈ
-
کلاؤڈ ڈیش بورڈ
-
ویب UI
-
کلاؤڈ ڈیش بورڈ
-
ہب ویب UI
-
نہیں
-
نہیں
-
نہیں
-
نہیں
-
-
-
سیلف-ہوسٹڈ
-
ہاں (ڈیفالٹ)
-
اختیاری
-
اختیاری
-
ہاں
-
نہیں (صرف کلاؤڈ)
-
ہاں (Docker)
-
ہاں
-
ہاں (Oracle DB)
-
ہاں
-
ہاں
-
-
-
-بینچ مارک نوٹ: صرف agentmemory کا R@5 ہمارا اپنا پیمائش شدہ نتیجہ ہے (LongMemEval-S، benchmark/COMPARISON.md سے دوبارہ قابلِ تولید)۔ mem0 اور Letta کے اعداد و شمار ان کے شائع شدہ LoCoMo نمبرز ہیں (ایک مختلف dataset)؛ MemPalace، supermemory، TencentDB (PersonaMem)، اور oracleagentmemory کے اعداد vendor کے خود رپورٹ شدہ دعوے ہیں جنہیں ہم نے آزادانہ طور پر دوبارہ تولید نہیں کیا (oracleagentmemory کا رن Oracle AI Database کے خلاف GPT-5.5 استعمال کرتا تھا)۔ محض اندازاً موازنے کے لیے ساتھ ساتھ دکھایا گیا ہے، ایک جیسے ڈیٹا پر براہِ راست مقابلہ نہیں۔ Star کی تعداد تخمینی ہے اور وقت کے ساتھ بدلتی رہتی ہے۔
-
-**نئے داخل ہونے والے** جن کے بارے میں جاننا مفید ہے، تفصیلی موازنہ [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) میں:
-
-| سسٹم | ⭐ | زاویہ |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Temporal نالج گراف؛ شائع شدہ temporal-query نتائج میں سب سے مضبوط (LongMemEval 63.8%)، لیکن گراف asynchronously بنتا ہے لہٰذا تازہ facts تاخیر سے آ سکتے ہیں |
-| Cognee | 30K | Document-to-knowledge-graph ingestion، صرف Python، structured entity extraction کے لیے بنایا گیا ہے، سیشن کیپچر کے لیے نہیں |
+مصنوعات اور تنصیب کے طریقوں کا تازہ موازنہ، حوالوں کے ساتھ:
-AgentMemory سیشن capture، قابل تلاش منصوبے کی تاریخ اور مقامی ویوئر اکٹھے فراہم کرتا ہے، اور پہلی بار سے بغیر کلید کام کر سکتا ہے۔
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -782,7 +574,7 @@ Verify with `curl http://localhost:3111/agentmemory/health`. Open http://localho
میموری سرور شروع کریں: `npx -y @agentmemory/agentmemory@latest`
-#### `npx skills add` کے ذریعے نیٹیو اسکلز (50+ ایجنٹس)
+#### `npx skills add` کے ذریعے نیٹیو اسکلز
agentmemory 17 اسکلز کو Claude-Code طرز کے `/SKILL.md` فارمیٹ میں شپ کرتا ہے: 9 invocable ایکشن اسکلز (`remember`، `recall`، `recap`، `handoff`، `forget`، `lesson`، `commit-context`، `commit-history`، `session-history`) اور 8 reference اسکلز جنہیں ایجنٹ ضرورت پڑنے پر لوڈ کرتا ہے (`memory-discipline`، `agentmemory-mcp-tools`، `agentmemory-rest-api`، `agentmemory-config`، `agentmemory-agents`، `agentmemory-hooks`، `agentmemory-architecture`، `write-agentmemory-skill`)۔ reference اسکلز سورس سے تیار کردہ data tables رکھتی ہیں، لہٰذا یہ کبھی drift نہیں کرتیں۔ vercel-labs کا [`skills`](https://npmjs.com/package/skills) CLI انہیں calling ایجنٹ کی نیٹیو اسکل ڈائرکٹری میں 50+ ایجنٹس (Claude Code، Cursor، Cline، Continue، Droid، Warp، Codex، Antigravity، Kiro، OpenCode، Goose، Roo، Trae، Windsurf، اور مزید) میں خودکار طور پر انسٹال کرتا ہے:
@@ -839,7 +631,7 @@ agentmemory کا entry ہر اس host پر **ایک ہی MCP سرور بلاک**
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` مشترکہ customization ڈائرکٹری میں MCP اور capture hooks انسٹال کرتا ہے۔ دیکھیں [Antigravity سیٹ اپ اور حدود](../docs/plugins/antigravity.md)۔ |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` موجودہ IDE ورژنز جیسی ہی MCP اور hook configuration استعمال کرتا ہے۔ موجودہ انسٹالیشنز کو `--force` کے ساتھ ریفریش کرنا چاہیے؛ دیکھیں [اپ گریڈ نوٹس](../docs/plugins/antigravity.md)۔ |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` user-level کنفیگ لکھتا ہے۔ Workspace overrides آپ کے کوڈ کے ساتھ `.kiro/settings/mcp.json` میں جاتے ہیں۔ |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` معیاری `mcpServers` بلاک لکھتا ہے۔ Warp `.claude/skills/` سے اسکلز کو بھی خودکار طور پر دریافت کرتا ہے؛ ایک بار Claude Code plugin انسٹال ہونے پر 8 agentmemory اسکلز (`remember`، `recall`، `recap`، `handoff`، `forget`، `commit-context`، `commit-history`، `session-history`) Warp کے slash-command palette میں نیٹیو طور پر ظاہر ہوتی ہیں۔ |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` معیاری `mcpServers` بلاک لکھتا ہے۔ Warp `.claude/skills/` سے اسکلز کو بھی خودکار طور پر دریافت کرتا ہے؛ ایک بار Claude Code plugin انسٹال ہونے پر 9 agentmemory اسکلز (`remember`، `recall`، `recap`، `handoff`، `forget`، `commit-context`، `commit-history`، `session-history`) Warp کے slash-command palette میں نیٹیو طور پر ظاہر ہوتی ہیں۔ |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` معیاری `mcpServers` بلاک لکھتا ہے۔ VS Code extension صارفین: وہی بلاک Cline Settings → MCP Servers → Edit JSON کے ذریعے پیسٹ کریں۔ |
| **Continue.dev** | `~/.continue/config.yaml` (ترجیحی) یا `config.json` (legacy) | `agentmemory connect continue` ترتیب نہ ہونے پر `config.yaml` بناتا ہے یا موجودہ `config.json` اپ ڈیٹ کرتا ہے۔ موجودہ YAML کے comments اور anchors محفوظ رکھنے کے لیے اسے تبدیل نہیں کرتا، بلکہ `mcpServers:` کے نیچے چسپاں کرنے کا بلاک دکھاتا ہے۔ Continue کا `mcpServers` ایک array ہے۔ |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` `context_servers` کے تحت لکھتا ہے (Zed کی اپنی key، `mcpServers` نہیں)۔ ریموٹ MCP سرورز کو بجائے `{"url": "..."}` کے ذریعے جوڑا جا سکتا ہے۔ |
@@ -930,7 +722,6 @@ Session 2: "Now add rate limiting"
|---|---|---|
| دیکھ بھال | ہاتھ سے ترتیب دی گئی ہدایات | انڈیکس شدہ سیشن کی تاریخ اور میموری |
| سرچ | سب کچھ context میں لوڈ کرتا ہے | BM25 + ویکٹر + گراف (صرف top-K) |
-| ٹوکن لاگت | 240 observations پر 22K+ | ~1,900 ٹوکنز (92% کم) |
| کراس-ایجنٹ | فی-ایجنٹ فائلیں | MCP + REST (کوئی بھی ایجنٹ) |
| کوآرڈینیشن | کوئی نہیں | Leases، سگنلز، actions، routines |
| Observability | دستی طور پر فائلیں پڑھنا | :3113 پر ریئل-ٹائم ویور |
@@ -943,7 +734,7 @@ Session 2: "Now add rate limiting"
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1028,7 +819,7 @@ Reciprocal Rank Fusion (RRF، k=60) کے ساتھ فیوز کیا گیا اور
جب ویکٹر index پاپولیٹڈ ہو، `mem::search` (`memory_recall` کے پیچھے) hybrid BM25 + vector ranker استعمال کرتا ہے۔ embeddings کے بغیر یہ BM25 استعمال کرتا ہے۔ `smart-search` گراف ڈیٹا موجود ہونے پر اضافی طور پر structural graph matches بھی فیوز کر سکتا ہے۔ Lesson recall ہر کوئری پر پورے corpus کو اسکین کرنے کے بجائے ایک وقف شدہ in-memory BM25 index پر چلتا ہے۔ سپرسیڈ شدہ میموری ورژنز ہر recall راستے سے خارج ہوتے ہیں؛ version chain ان کی ہسٹری رکھتی ہے۔
-ویکٹرز کریش یا force-kill سے بچ جاتے ہیں۔ ویکٹر index کم از کم ہر `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 منٹ) بعد buckets میں محفوظ ہوتا ہے۔ اس دوران شامل یا ہٹایا گیا ہر ویکٹر state store میں ایک چھوٹے pending log میں فوراً لکھ دیا جاتا ہے، اور اگلا اسٹارٹ اسے embedding provider کو کال کیے بغیر replay کرتا ہے۔ ہر کامیاب save لاگ کو خالی کر دیتا ہے۔ وہ documents جن کا replay کے بعد بھی کوئی ویکٹر نہیں ہوتا انہیں پسِ منظر میں `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) کے batches میں دوبارہ embed کیا جاتا ہے جب تک کوئی باقی نہ رہے، اور رکا ہوا backfill اگلے اسٹارٹ پر جاری رہتا ہے۔ `/agentmemory/status` اور ویور pending log کا سائز اور backfill اسٹیٹس دکھاتے ہیں۔ Keyless انسٹالز کچھ نہیں لکھتے۔
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 بذاتِ خود یونانی، سیریلک، عبرانی، عربی، اور accented Latin کو ٹوکنائز کرتا ہے۔ چینی / جاپانی / کوریائی میموریز کے لیے، CJK runs کو word-level tokens میں تقسیم کرنے کے لیے اختیاری segmenters (`npm install @node-rs/jieba tiny-segmenter`) انسٹال کریں؛ ان کے بغیر، agentmemory نرمی سے whole-run tokenization پر fall ہو جاتا ہے اور stderr پر ایک بار اشارہ پرنٹ کرتا ہے۔
@@ -1044,9 +835,9 @@ EMBEDDING_PROVIDER=local
| Provider | ماڈل | لاگت | نوٹس |
|---|---|---|---|
-| **Local (تجویز کردہ opt-in)** | `all-MiniLM-L6-v2` | مفت | پہلے ماڈل ڈاؤن لوڈ کے بعد آن-ڈیوائس، صرف-BM25 کے مقابلے میں +8pp recall |
+| **Local (تجویز کردہ opt-in)** | `all-MiniLM-L6-v2` | مفت | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | مفت ٹیئر | 100+ زبانیں، 768/1536/3072 dims (MRL)، 2048-ٹوکن input۔ `text-embedding-004` کی جگہ لیتا ہے ([deprecated، shutdown Jan 14, 2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | بہترین کوالٹی |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | ادا شدہ (Paid) | کوڈ کے لیے optimized |
| Cohere | `embed-english-v3.0` | مفت trial | عمومی مقصد |
| OpenRouter | کوئی بھی ماڈل | مختلف ہوتا ہے | Multi-model پراکسی |
diff --git a/READMEs/README.vi-VN.md b/READMEs/README.vi-VN.md
index 06aea9a278..6a80a5fe14 100644
--- a/READMEs/README.vi-VN.md
+++ b/READMEs/README.vi-VN.md
@@ -66,8 +66,6 @@
-
-
@@ -331,223 +329,17 @@ Ghi chú phát hành mới nhất: [CHANGELOG.md](../CHANGELOG.md).
-
-
-
-
-### Độ chính xác Retrieval
-
-**coding-agent-life-v1** (corpus nội bộ, có thể tái lập trong sandbox)
-
-| Adapter | P@5 | R@5 | Tỷ lệ hit top-5 | Độ trễ p50 |
-|---|---|---|---|---|
-| **agentmemory hybrid** | **0.240** | **1.000** | **15 / 15** | 14 ms |
-| grep baseline | 0.227 | 0.967 | 15 / 15 | 0 ms |
-
-Tỷ lệ hit top-5 đạt 100% tại **ngưỡng toán học P@5** cho corpus này (0.240, xem scorecard). Hybrid lấy được mọi gold session; grep bỏ lỡ 1 trong 2 gold ở câu truy vấn temporal đa session. Mức cải thiện nằm ở **recall + temporal**, không phải precision tổng hợp. Benchmark này nhỏ và thưa gold; LongMemEval-S lớn hơn ở dưới phân biệt tốt hơn. Phân tích chi tiết theo từng loại + ghi chú đính chính đầy đủ: [`docs/benchmarks/2026-05-20-coding-agent-life-v1.md`](../docs/benchmarks/2026-05-20-coding-agent-life-v1.md).
-
-**LongMemEval-S** (ICLR 2025, 500 câu hỏi)
+Kết quả truy xuất hiện tại, định nghĩa chỉ số và các bước tái lập:
-| Hệ thống | R@5 | R@10 | MRR |
-|---|---|---|---|
-| **agentmemory** | **95.2%** | **98.6%** | **88.2%** |
-| BM25-only fallback | 86.2% | 94.6% | 71.5% |
-
-
-
-
-### Tiết kiệm Token
-
-| Cách tiếp cận | Token/năm | Chi phí/năm |
-|---|---|---|
-| Dán toàn bộ context | 19.5M+ | Không thể (vượt window) |
-| Tóm tắt bằng LLM | ~650K | ~$500 |
-| **agentmemory** | **~170K** | **~$10** |
-| agentmemory + embedding cục bộ | ~170K | **$0** |
-
-
-
-
-
-> Embedding model: `all-MiniLM-L6-v2` (cục bộ, miễn phí, không cần API key). Báo cáo đầy đủ: [`benchmark/LONGMEMEVAL.md`](../benchmark/LONGMEMEVAL.md), [`benchmark/QUALITY.md`](../benchmark/QUALITY.md), [`benchmark/SCALE.md`](../benchmark/SCALE.md). So sánh đối thủ: [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md) bao gồm agentmemory so với mem0, Letta, Khoj, supermemory, TencentDB Agent Memory, MemPalace, Zep/Graphiti, Cognee, Hippo.
-
-**Tái lập cục bộ:** [`eval/README.md`](../eval/README.md), một harness có thể cắm (pluggable) adapter cho LongMemEval `_s` (500 câu hỏi công khai) + `coding-agent-life-v1` (corpus nội bộ 15 session). Các adapter grep / vector / agentmemory được chấm điểm song song, xuất NDJSON, các scorecard đã công bố nằm tại [`docs/benchmarks/`](../docs/benchmarks/).
-
-**Kết hợp tốt với [codegraph](https://github.com/colbymchenry/codegraph), [Understand Anything](https://github.com/Lum1104/Understand-Anything), và [Graphify](https://github.com/safishamsi/graphify).** Index hóa code-graph, pipeline build đa agent, và các knowledge graph rộng hơn trên docs / PDF / ảnh / video. agentmemory ghi nhớ phần công việc; ba dự án này làm sáng rõ phần còn lại của context layer. Công thức + bảng định tuyến câu hỏi: [`docs/recipes/pairings.md`](../docs/recipes/pairings.md).
+[0.9.31 · 2026-10-28 · strict recall-all@5](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) · [eval/README.md](../eval/README.md).
---
-
-
-
-
agentmemory
-
mem0 (63K ⭐)
-
Letta / MemGPT (24K ⭐)
-
Khoj (36K ⭐)
-
supermemory (29K ⭐)
-
TencentDB Agent Memory (22K ⭐)
-
MemPalace (54K ⭐)
-
oracleagentmemory
-
Hippo
-
Tích hợp sẵn (CLAUDE.md)
-
-
-
Loại
-
Engine bộ nhớ + MCP server
-
API lớp bộ nhớ
-
Runtime agent đầy đủ
-
AI cá nhân
-
API bộ nhớ + app
-
Hub bộ nhớ nhóm (LLM proxy)
-
Bộ nhớ vector (OSS)
-
Engine bộ nhớ (Oracle DB)
-
Hệ thống bộ nhớ
-
File tĩnh
-
-
-
Retrieval R@5
-
95.2%
-
68.5% (LoCoMo)
-
83.2% (LoCoMo)
-
N/A
-
Tự công bố
-
PersonaMem 76% (tự công bố)
-
~96.6% (tự công bố)
-
94.4% (tự công bố)
-
N/A
-
N/A (grep)
-
-
-
Auto-capture
-
12 hooks (không cần thao tác thủ công)
-
Gọi add() thủ công
-
Agent tự sửa
-
Thủ công
-
Trích xuất phía API
-
Chặn qua proxy (đổi base-URL)
-
Thủ công
-
Trích xuất qua API
-
Thủ công
-
Chỉnh sửa thủ công
-
-
-
Tìm kiếm
-
BM25 + Vector + Graph (RRF fusion)
-
Vector + Graph
-
Vector (lưu trữ archival)
-
Ngữ nghĩa (semantic)
-
Vector + RAG
-
4 loại asset (Chat / Skill / Wiki / CodeGraph)
-
Chỉ Vector
-
Vector + ngữ nghĩa
-
Có trọng số suy giảm (decay)
-
Nạp tất cả vào context
-
-
-
Đa agent
-
MCP + REST + lease + signal
-
API (không điều phối)
-
Chỉ trong runtime Letta
-
Không
-
Không
-
Vai trò nhóm + asset chia sẻ
-
Không
-
Chỉ theo scope
-
Chia sẻ đa agent
-
File theo từng agent
-
-
-
Lệ thuộc framework
-
Không (bất kỳ MCP client nào)
-
Không
-
Cao (phải dùng Letta)
-
Độc lập (standalone)
-
Không
-
Proxy chặn trước mọi lệnh gọi model
-
Không
-
Oracle Database
-
Không
-
Định dạng theo từng agent
-
-
-
Dependency bên ngoài
-
Không cần dịch vụ cơ sở dữ liệu riêng
-
Qdrant / pgvector
-
Postgres + vector DB
-
Nhiều
-
Cloud được quản lý
-
Docker stack (Core + Hub + Proxy)
-
Vector store
-
Oracle AI Database
-
Không
-
Không
-
-
-
Lifecycle bộ nhớ
-
Hợp nhất 4 tầng + suy giảm + tự quên
-
Trích xuất bị động
-
Do agent quản lý
-
Thủ công
-
Tự quên (auto-forget)
-
Review thủ công; auto-routing đang phát triển
-
Không
-
Không công bố
-
Suy giảm + hợp nhất
-
Cắt bỏ thủ công
-
-
-
Hiệu quả token
-
~1,900 token/session ($10/năm)
-
Thay đổi theo tích hợp
-
Bộ nhớ lõi (core memory) nằm trong context
-
Thay đổi
-
Theo giá cloud
-
Không công bố
-
Không có ngân sách token
-
Dựa trên LLM (thay đổi)
-
Thay đổi
-
22K+ token ở 240 observation
-
-
-
Viewer real-time
-
Có (port 3113)
-
Cloud dashboard
-
Cloud dashboard
-
Web UI
-
Cloud dashboard
-
Hub web UI
-
Không
-
Không
-
Không
-
Không
-
-
-
Self-hosted
-
Có (mặc định)
-
Tùy chọn
-
Tùy chọn
-
Có
-
Không (chỉ cloud)
-
Có (Docker)
-
Có
-
Có (Oracle DB)
-
Có
-
Có
-
-
-
-Ghi chú benchmark: chỉ con số R@5 của agentmemory là kết quả đo đạc của chính chúng tôi (LongMemEval-S, có thể tái lập từ benchmark/COMPARISON.md). Các số liệu của mem0 và Letta là số LoCoMo họ đã công bố (một dataset khác); các số liệu của MemPalace, supermemory, TencentDB (PersonaMem), và oracleagentmemory là các tuyên bố tự công bố từ nhà cung cấp mà chúng tôi chưa tái lập độc lập (lần chạy của oracleagentmemory dùng GPT-5.5 trên một Oracle AI Database). Chỉ hiển thị song song để ước lượng, không phải so sánh trực tiếp (head-to-head) trên cùng dữ liệu. Số lượng star chỉ là gần đúng và thay đổi theo thời gian.
-
-**Những cái tên mới** đáng để biết, được so sánh chi tiết trong [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| Hệ thống | ⭐ | Hướng tiếp cận |
-|--------|---|-------|
-| Zep / Graphiti | 30K | Knowledge graph theo thời gian (temporal); kết quả truy vấn temporal công bố mạnh nhất (LongMemEval 63.8%), nhưng graph được xây dựng bất đồng bộ nên các fact mới có thể bị trễ |
-| Cognee | 30K | Nạp từ tài liệu sang knowledge graph, chỉ hỗ trợ Python, được xây dựng để trích xuất entity có cấu trúc hơn là để capture session |
+So sánh cập nhật về sản phẩm và phương án triển khai, kèm nguồn:
-AgentMemory kết hợp ghi nhận phiên, lịch sử dự án có thể tìm kiếm và trình xem cục bộ, với chế độ không cần khóa ngay từ lần đầu.
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -780,7 +572,7 @@ Hướng dẫn đầy đủ: [`integrations/hermes/`](../integrations/hermes/)
Khởi động memory server: `npx -y @agentmemory/agentmemory@latest`
-#### Skill gốc qua `npx skills add` (50+ agent)
+#### Skill gốc qua `npx skills add`
agentmemory đi kèm 17 skill theo định dạng `/SKILL.md` kiểu Claude Code: 9 skill hành động có thể gọi (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) và 8 skill tham chiếu mà agent nạp theo nhu cầu (`memory-discipline`, `agentmemory-mcp-tools`, `agentmemory-rest-api`, `agentmemory-config`, `agentmemory-agents`, `agentmemory-hooks`, `agentmemory-architecture`, `write-agentmemory-skill`). Các skill tham chiếu mang theo bảng dữ liệu được sinh ra từ source, nên chúng không bao giờ lệch (drift). CLI [`skills`](https://npmjs.com/package/skills) của vercel-labs tự cài chúng vào thư mục skill gốc của agent gọi nó, trên hơn 50 agent (Claude Code, Cursor, Cline, Continue, Droid, Warp, Codex, Antigravity, Kiro, OpenCode, Goose, Roo, Trae, Windsurf, và nhiều hơn nữa):
@@ -837,7 +629,7 @@ Entry của agentmemory là **cùng một block MCP server** trên mọi host d
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` cài MCP và capture hook vào thư mục customization chung. Xem [thiết lập và giới hạn của Antigravity](../docs/plugins/antigravity.md). |
| **Antigravity CLI** (`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` dùng cùng cấu hình MCP và hook như các phiên bản IDE hiện tại. Các bản cài đặt sẵn có nên refresh bằng `--force`; xem [ghi chú upgrade](../docs/plugins/antigravity.md). |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` viết config ở cấp user. Các override ở cấp workspace nằm trong `.kiro/settings/mcp.json` cạnh code của bạn. |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` viết block `mcpServers` chuẩn. Warp cũng tự phát hiện skill từ `.claude/skills/`; khi plugin Claude Code được cài, 8 skill agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `commit-context`, `commit-history`, `session-history`) sẽ xuất hiện một cách tự nhiên trong bảng slash-command của Warp. |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` viết block `mcpServers` chuẩn. Warp cũng tự phát hiện skill từ `.claude/skills/`; khi plugin Claude Code được cài, 9 skill agentmemory (`remember`, `recall`, `recap`, `handoff`, `forget`, `lesson`, `commit-context`, `commit-history`, `session-history`) sẽ xuất hiện một cách tự nhiên trong bảng slash-command của Warp. |
| **Cline (CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` viết block `mcpServers` chuẩn. Người dùng extension VS Code: dán cùng block này qua Cline Settings → MCP Servers → Edit JSON. |
| **Continue.dev** | `~/.continue/config.yaml` (ưu tiên) hoặc `config.json` (legacy) | `agentmemory connect continue` tạo `config.yaml` nếu chưa có cấu hình hoặc cập nhật `config.json` hiện có. YAML có sẵn được giữ nguyên để bảo toàn chú thích và anchor; lệnh in khối để dán dưới `mcpServers:`. Continue dùng dạng mảng cho `mcpServers`. |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` viết dưới `context_servers` (key riêng của Zed, KHÔNG phải `mcpServers`). Các MCP server remote có thể được kết nối qua `{"url": "..."}` thay thế. |
@@ -933,7 +725,6 @@ Mọi AI coding agent đều đi kèm bộ nhớ tích hợp sẵn: Claude Code
|---|---|---|
| Bảo trì | Chỉ dẫn được quản lý thủ công | Lịch sử phiên và bộ nhớ đã lập chỉ mục |
| Tìm kiếm | Nạp tất cả vào context | BM25 + vector + graph (chỉ top-K) |
-| Chi phí token | 22K+ ở 240 observation | ~1,900 token (giảm 92%) |
| Đa agent | File theo từng agent | MCP + REST (mọi agent) |
| Điều phối | Không có | Lease, signal, action, routine |
| Observability | Đọc file thủ công | Viewer real-time trên :3113 |
@@ -946,7 +737,7 @@ Mọi AI coding agent đều đi kèm bộ nhớ tích hợp sẵn: Claude Code
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1031,7 +822,7 @@ Retrieval ba luồng (triple-stream) kết hợp ba tín hiệu:
Khi vector index đã có dữ liệu, `mem::search` (đứng sau `memory_recall`) dùng ranker hybrid BM25 + vector. Khi không có embedding, nó dùng BM25. `smart-search` còn có thể fuse thêm các kết quả khớp từ graph cấu trúc khi dữ liệu graph tồn tại, kể cả ở chế độ keyless. Recall cho lesson chạy trên một BM25 index riêng trong memory thay vì quét toàn bộ corpus cho mỗi truy vấn. Các phiên bản memory đã bị supersede bị loại khỏi mọi đường recall; chuỗi version vẫn giữ lại lịch sử của chúng.
-Vector sống sót qua một lần crash hoặc force-kill. Vector index được lưu theo bucket, nhiều nhất là mỗi `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS` (10 phút). Mỗi vector được thêm hoặc xóa ở giữa cũng được ghi ngay vào một pending log nhỏ trong state store, và lần khởi động kế tiếp sẽ replay nó mà không cần gọi embedding provider. Mỗi lần save thành công sẽ làm trống log đó. Các document vẫn chưa có vector sau khi replay sẽ được re-embed ở nền theo từng batch `AGENTMEMORY_VECTOR_BACKFILL_MAX` (500) cho tới khi không còn document nào, và một backfill bị dừng sẽ tiếp tục ở lần khởi động kế tiếp. `/agentmemory/status` và viewer hiển thị kích thước pending log và trạng thái backfill. Bản cài keyless không ghi gì cả.
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 tokenize được tiếng Hy Lạp, Cyrillic, Hebrew, Ả Rập, và Latin có dấu ngay từ đầu. Với bộ nhớ tiếng Trung / Nhật / Hàn, hãy cài thêm các segmenter tùy chọn (`npm install @node-rs/jieba tiny-segmenter`) để chia các đoạn CJK thành token ở mức từ; nếu không có chúng, agentmemory sẽ lùi nhẹ về tokenize theo cả đoạn và in ra một gợi ý một lần trên stderr.
@@ -1047,9 +838,9 @@ Lần cài npm thông thường đã bao gồm runtime tùy chọn `@huggingface
| Provider | Model | Chi phí | Ghi chú |
|---|---|---|---|
-| **Local (khuyến nghị opt-in)** | `all-MiniLM-L6-v2` | Miễn phí | Chạy trên máy sau lần tải model đầu tiên, +8pp recall so với chỉ dùng BM25 |
+| **Local (khuyến nghị opt-in)** | `all-MiniLM-L6-v2` | Miễn phí | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | Free tier | 100+ ngôn ngữ, 768/1536/3072 chiều (MRL), input 2048 token. Thay thế `text-embedding-004` ([đã deprecated, shutdown 14/01/2026](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | Chất lượng cao nhất |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | Trả phí | Tối ưu cho code |
| Cohere | `embed-english-v3.0` | Free trial | Dùng chung (general purpose) |
| OpenRouter | Bất kỳ model nào | Thay đổi | Proxy đa model |
diff --git a/READMEs/README.zh-CN.md b/READMEs/README.zh-CN.md
index 0941ab0d5d..68459211ea 100644
--- a/READMEs/README.zh-CN.md
+++ b/READMEs/README.zh-CN.md
@@ -66,8 +66,6 @@
-
-基準測試說明:只有 agentmemory 的 R@5 是我們自己量測的結果(LongMemEval-S,可從 benchmark/COMPARISON.md 重現)。mem0 和 Letta 的數字是它們公開發表的 LoCoMo 數據(屬於不同的資料集);MemPalace、supermemory、TencentDB(PersonaMem)和 oracleagentmemory 的數字是廠商自行回報、我們尚未獨立重現的宣稱(oracleagentmemory 的測試使用 GPT-5.5 對上 Oracle AI Database)。並列呈現僅供大致參考,並非在相同資料上的正面對比。星數為近似值,會隨時間變動。
-
-值得關注的**新進者**,詳細比較見 [`benchmark/COMPARISON.md`](../benchmark/COMPARISON.md):
-
-| 系統 | ⭐ | 切入點 |
-|--------|---|-------|
-| Zep / Graphiti | 30K | 時間性知識圖譜;已發表的時間性查詢結果最強(LongMemEval 63.8%),但圖是非同步建構,因此最新事實可能有延遲 |
-| Cognee | 30K | 文件到知識圖譜的擷取,僅支援 Python,專為結構化實體擷取而設計,而非會話捕捉 |
+產品與部署方式的最新比較及來源:
-AgentMemory 結合工作階段捕捉、可搜尋的專案歷史和本機檢視器,首次使用即可選擇免金鑰模式。
+[AgentMemory / Mem0 / Supermemory / Zep / Letta / Cognee / claude-mem](../benchmark/COMPARISON.md) · 2026-10-28.
---
@@ -778,7 +570,7 @@ Verify with `curl http://localhost:3111/agentmemory/health`. Open http://localho
啟動記憶伺服器:`npx -y @agentmemory/agentmemory@latest`
-#### 透過 `npx skills add` 使用原生 skills(50+ 代理)
+#### 透過 `npx skills add` 使用原生 skills
agentmemory 以 Claude-Code 風格的 `/SKILL.md` 格式提供 17 個 skills:9 個可呼叫的動作 skills(`remember`、`recall`、`recap`、`handoff`、`forget`、`lesson`、`commit-context`、`commit-history`、`session-history`)和 8 個代理按需載入的參考 skills(`memory-discipline`、`agentmemory-mcp-tools`、`agentmemory-rest-api`、`agentmemory-config`、`agentmemory-agents`、`agentmemory-hooks`、`agentmemory-architecture`、`write-agentmemory-skill`)。這些參考 skills 內含從原始碼產生的資料表,因此永遠不會漂移。vercel-labs 的 [`skills`](https://npmjs.com/package/skills) CLI 會把它們自動安裝到發起代理的原生 skill 目錄中,支援 50+ 個代理(Claude Code、Cursor、Cline、Continue、Droid、Warp、Codex、Antigravity、Kiro、OpenCode、Goose、Roo、Trae、Windsurf 等):
@@ -835,7 +627,7 @@ Hermes、OpenClaw、Pi 和 Oh My Pi 的連接器預設安裝原生整合和 skil
| **Antigravity IDE / 2.0** | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity --with-hooks` 會在共用的自訂目錄中安裝 MCP 和擷取 hooks。參見[Antigravity 設定與限制](../docs/plugins/antigravity.md)。 |
| **Antigravity CLI**(`agy`) | `~/.gemini/config/mcp_config.json` | `agentmemory connect antigravity-cli --with-hooks` 使用與目前 IDE 版本相同的 MCP 和 hook 設定。既有安裝應使用 `--force` 刷新;參見[升級說明](../docs/plugins/antigravity.md)。 |
| **Kiro** | `~/.kiro/settings/mcp.json` | `agentmemory connect kiro` 會寫入使用者層級的設定。工作區層級的覆寫請放在你程式碼旁的 `.kiro/settings/mcp.json`。 |
-| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` 會寫入標準的 `mcpServers` 區塊。Warp 也會自動從 `.claude/skills/` 探索 skills;一旦安裝了 Claude Code 外掛,8 個 agentmemory skills(`remember`、`recall`、`recap`、`handoff`、`forget`、`commit-context`、`commit-history`、`session-history`)就會原生出現在 Warp 的斜線指令選單中。 |
+| **Warp** | `~/.warp/.mcp.json` | `agentmemory connect warp` 會寫入標準的 `mcpServers` 區塊。Warp 也會自動從 `.claude/skills/` 探索 skills;一旦安裝了 Claude Code 外掛,9 個 agentmemory skills(`remember`、`recall`、`recap`、`handoff`、`forget`、`commit-context`、`commit-history`、`session-history`)就會原生出現在 Warp 的斜線指令選單中。 |
| **Cline(CLI)** | `~/.cline/mcp.json` | `agentmemory connect cline` 會寫入標準的 `mcpServers` 區塊。VS Code 擴充功能使用者:透過 Cline Settings → MCP Servers → Edit JSON 貼上同一個區塊。 |
| **Continue.dev** | `~/.continue/config.yaml`(建議)或 `config.json`(舊版) | `agentmemory connect continue` 在沒有設定時建立 `config.yaml`,或更新既有 `config.json`。既有 YAML 會保留不動,以保護註解和錨點;指令會印出應貼到 `mcpServers:` 下的設定區塊。Continue 的 `mcpServers` 使用陣列。 |
| **Zed** | `~/.config/zed/settings.json` | `agentmemory connect zed` 會寫入 `context_servers`(Zed 自己的鍵,**不是** `mcpServers`)。遠端 MCP 伺服器可改用 `{"url": "..."}` 接上。 |
@@ -926,7 +718,6 @@ Session 2: "Now add rate limiting"
|---|---|---|
| 維護 | 手動整理的指令 | 已索引的工作階段歷史與記憶 |
| 搜尋 | 把所有內容載入上下文 | BM25 + 向量 + 圖(僅 top-K) |
-| Token 成本 | 240 條觀測達 22K+ | ~1,900 個 token(少 92%) |
| 跨代理 | 每個代理各自的檔案 | MCP + REST(任何代理) |
| 協調 | 無 | 租約、訊號、動作、例程 |
| 可觀測性 | 手動讀檔 | 連接埠 3113 的即時檢視器 |
@@ -939,7 +730,7 @@ Session 2: "Now add rate limiting"
```text
PostToolUse hook fires
- -> SHA-256 dedup (5min window)
+ -> Stable event ID / durable capture inbox / offline spool
-> Privacy filter (strip secrets, API keys)
-> Store raw observation
-> Synthetic compression by default
@@ -1024,7 +815,7 @@ SessionStart hook fires
當向量索引已有資料時,`mem::search`(`memory_recall` 背後)會使用混合的 BM25 + 向量排序器。沒有嵌入時則使用 BM25。當圖資料存在時,`smart-search` 還能額外融合結構化的圖比對結果,即使在無金鑰模式下也是如此。教訓召回在專用的記憶體內 BM25 索引上執行,而非每次查詢都掃描整個語料庫。被取代的記憶版本會從每一條召回路徑中排除;版本鏈保留它們的歷史。
-向量可在崩潰或強制終止後存活。向量索引最多每 `AGENTMEMORY_INDEX_SAVE_INTERVAL_MS`(10 分鐘)以分桶方式儲存一次。期間新增或移除的每個向量,也會立即寫入狀態儲存中的一份小型待處理日誌,下次啟動時會重播它而不呼叫嵌入提供者。每次成功儲存都會清空該日誌。重播後仍沒有向量的文件,會在背景以每批 `AGENTMEMORY_VECTOR_BACKFILL_MAX`(500)筆的方式重新嵌入,直到沒有剩餘為止;被中止的回填會在下次啟動時繼續。`/agentmemory/status` 和檢視器會顯示待處理日誌大小與回填狀態。無金鑰安裝則完全不寫入任何內容。
+[0.9.31: `AGENTMEMORY_VECTOR_BACKFILL=all` / `AGENTMEMORY_VECTOR_BACKFILL_MAX`](../docs/reliability/source-and-vector-recovery.md).
BM25 開箱即可對希臘文、西里爾文、希伯來文、阿拉伯文和帶重音的拉丁文進行分詞。對於中文 / 日文 / 韓文的記憶,安裝選用的分詞器(`npm install @node-rs/jieba tiny-segmenter`),把 CJK 文字段切成詞級 token;若未安裝,agentmemory 會優雅地退回整段分詞,並在 stderr 上印出一次性提示。
@@ -1040,9 +831,9 @@ EMBEDDING_PROVIDER=local
| 提供者 | 模型 | 成本 | 備註 |
|---|---|---|---|
-| **本地(建議選用)** | `all-MiniLM-L6-v2` | 免費 | 第一次模型下載後即為裝置端運算,召回率比僅用 BM25 高 +8pp |
+| **本地(建議選用)** | `all-MiniLM-L6-v2` | 免費 | [`EMBEDDING_PROVIDER=local`](../README.md#embedding-providers) |
| Gemini | `gemini-embedding-001` | 免費額度 | 支援 100+ 種語言,768/1536/3072 維(MRL),輸入上限 2048 token。取代已棄用的 `text-embedding-004`([2026 年 1 月 14 日停用](https://ai.google.dev/gemini-api/docs/deprecations)) |
-| OpenAI | `text-embedding-3-small` | $0.02/1M | 品質最高 |
+| OpenAI | `text-embedding-3-small` | [OpenAI](https://openai.com/api/pricing/) | [text-embedding-3-small](https://platform.openai.com/docs/models/text-embedding-3-small) |
| Voyage AI | `voyage-code-3` | 付費 | 針對程式碼優化 |
| Cohere | `embed-english-v3.0` | 免費試用 | 通用型 |
| OpenRouter | 任何模型 | 依情況而異 | 多模型代理 |
diff --git a/benchmark/COMPARISON.md b/benchmark/COMPARISON.md
index 207ca0f685..23bdb749d3 100644
--- a/benchmark/COMPARISON.md
+++ b/benchmark/COMPARISON.md
@@ -1,181 +1,27 @@
-# AI Agent Memory: Benchmark Comparison
+# Choosing a memory system
-How agentmemory compares against other persistent memory solutions for AI coding agents.
+Checked against the linked primary sources on 2026-10-28. Products evolve; deployment, model usage and commercial plans depend on the selected configuration.
-All numbers here come from published benchmarks or public repositories. We link to primary sources wherever possible so you can reproduce.
+| Project | Main use | Deployment and integration |
+|---|---|---|
+| **[AgentMemory](../README.md)** | Shared project memory for existing coding agents | Apache-2.0 local service; automatic capture on supported integrations, MCP and REST. BM25 works without a model key; embeddings and LLM processing are optional. The CLI installs its runtime. |
+| **[Mem0](https://docs.mem0.ai/introduction)** | Memory APIs for applications and coding agents | Hosted platform or open-source library/server; SDKs, MCP and coding-agent plugins. Capture depends on the integration. [Platform and OSS differ](https://docs.mem0.ai/platform/platform-vs-oss), including graph availability. |
+| **[Supermemory](https://supermemory.ai/docs/self-hosting/overview)** | Memory and document context for applications and agents | Hosted API or local server binary, plus coding-agent plugins. The local server has embedded storage and local embeddings; its codebase is separate and non-public. Local and hosted capabilities differ. |
+| **[Zep / Graphiti](https://github.com/getzep/graphiti)** | Enterprise context and temporal knowledge graphs | Zep offers a managed or BYOC platform. Apache-2.0 Graphiti is a separate self-hosted graph framework with database and model configuration. |
+| **[Letta](https://docs.letta.com/)** | Stateful agents with managed memory | Agent harness, CLI and SDK with MemFS/shared memory; managed cloud or self-hosted runtime. It supplies the agent runtime as well as memory. |
+| **[Cognee](https://github.com/topoteretes/cognee)** | Searchable knowledge from documents, code and conversations | Open-source local library/server or Cognee Cloud, with plugins and MCP. Local extraction and embedding models support a keyless path. |
+| **[claude-mem](https://github.com/thedotmack/claude-mem)** | Capture and summarize coding sessions | Apache-2.0 project with multiple agent integrations. Model-backed summaries can use its hosted observer or a configured provider; installation and provider choices affect data flow and cost. |
----
+AgentMemory fits developers who want one local memory service shared by their existing tools, with inspectable observations, lessons, actions, source expansion and a live viewer. An application-facing API, managed enterprise platform, document graph or complete stateful-agent runtime may suit a different workload better.
-## Retrieval Accuracy (LongMemEval)
+## Compare measurements on the same terms
-[LongMemEval](https://arxiv.org/abs/2410.10813) (ICLR 2025) measures long-term memory retrieval across ~48 sessions per question on the S variant (500 questions, ~115K tokens each).
+The [current retrieval report](../docs/benchmarks/2026-10-28-retrieval-evaluation.md) measures the BM25 component and a grep baseline over identical supplied sessions. It includes strict recall-all@5, partial recall, any-hit, coverage and failures. It does not rank the products above.
-| System | Benchmark | R@5 | Notes |
-|---|---|---|---|
-| **agentmemory** (BM25 + Vector) | LongMemEval-S | **95.2%** | `all-MiniLM-L6-v2` embeddings, no API key |
-| agentmemory (BM25-only) | LongMemEval-S | 86.2% | Fallback when no embedding provider available |
-| MemPalace | LongMemEval-S | ~96.6% (self-reported) | Vendor-published number we have not independently reproduced. Vector-only with a larger embedding model and no agent-integration surface (no hooks, no MCP, no multi-agent) |
-| oracleagentmemory | LongMemEval | 94.4% (self-reported) | Vendor-published, scored with GPT-5.5 at "xhigh reasoning" and requires an Oracle AI Database. We have not reproduced it. agentmemory's 95.2% uses free local embeddings and no API key |
-| Letta / MemGPT | LoCoMo | 83.2% | Different benchmark (LoCoMo, not LongMemEval) |
-| Mem0 | LoCoMo | 68.5% | Different benchmark (LoCoMo, not LongMemEval) |
+A useful comparison must hold the dataset revision, ingestion path, retrieved unit, cutoff, model, context budget, reader and judge constant. Retrieval recall, answer accuracy, task completion and token use measure different things. Historical scores and unrelated datasets should not share a leaderboard.
-**⚠️ Apples vs oranges caveat:** only agentmemory's 95.2% is our own measured result, reproducible from the methodology below. Every other number here is the vendor's published claim, on a different benchmark or harness, that we have not independently reproduced: MemPalace and oracleagentmemory report LongMemEval (oracleagentmemory's run used GPT-5.5 at "xhigh reasoning" against an Oracle AI Database), while Letta and Mem0 publish on [LoCoMo](https://snap-stanford.github.io/LoCoMo/). Treat them as ballpark vendor claims, not a head-to-head on identical data. We'd love to run every system on the same dataset; if any maintainer wants to collaborate, open an issue.
+## Costs and data flow
-Full agentmemory methodology: [`LONGMEMEVAL.md`](LONGMEMEVAL.md)
+AgentMemory has no license fee or hosted subscription. You operate its local runtime; optional hosted models charge for their usage. Installation, updates and first-use local model downloads need network access. Memory text stays local with keyless search or local models; configured remote services receive the text they process.
----
-
-## Feature Matrix
-
-| Feature | agentmemory | mem0 | Letta/MemGPT | Khoj | supermemory | MemPalace | oracleagentmemory | Hippo |
-|---|---|---|---|---|---|---|---|---|
-| **GitHub stars** | Growing | 58K+ | 23K+ | 35K+ | 26K+ | 54K+ | PyPI (Oracle) | Trending |
-| **Type** | Memory engine + MCP server | Memory layer API | Full agent runtime | Personal AI | Memory API + app | Benchmark-focused OSS | Memory engine (Oracle DB) | Memory system |
-| **Auto-capture via hooks** | ✅ 12 lifecycle hooks | ❌ Manual `add()` | ❌ Agent self-edits | ❌ Manual | ❌ API-side extraction | ❌ Manual | ❌ API extraction | ❌ Manual |
-| **Search strategy** | BM25 + Vector + Graph | Vector + Graph | Vector (archival) | Semantic | Vector + RAG | Vector-only (large model) | Vector + semantic | Decay-weighted |
-| **Multi-agent coordination** | ✅ Leases + signals + mesh | ❌ | Runtime-internal only | ❌ | ❌ | ❌ | Scoped only (user/agent/thread) | Multi-agent shared |
-| **Framework lock-in** | None | None | High | Standalone | None (drop-in wrappers) | None | Oracle Database | None |
-| **External deps** | None | Qdrant/pgvector | Postgres + vector | Multiple | Managed cloud | Vector store | Oracle AI Database | None |
-| **Self-hostable** | ✅ default | Optional | Optional | ✅ | ❌ Cloud-only | ✅ | ✅ (needs Oracle DB) | ✅ |
-| **Knowledge graph** | ✅ Entity extraction + BFS | ✅ Mem0g variant | ❌ | Doc links | ❌ | ❌ | ❌ | ❌ |
-| **Memory decay** | ✅ Ebbinghaus + tiered | ❌ | ❌ | ❌ | ✅ Auto-forget | ❌ | ❌ | ✅ Half-lives |
-| **4-tier consolidation** | ✅ Working → episodic → semantic → procedural | ❌ | OS-inspired tiers | ❌ | ❌ | ❌ | ❌ | Episodic + semantic |
-| **Version / supersession** | ✅ Jaccard-based | Passive | ❌ | ❌ | ✅ Auto-resolve | ❌ | ❌ | ❌ |
-| **Real-time viewer** | ✅ Port 3113 | Cloud dashboard | Cloud dashboard | Web UI | Cloud dashboard | ❌ | ❌ | ❌ |
-| **Privacy filtering** | ✅ Strips secrets pre-store | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-| **Obsidian export** | ✅ Built-in | ❌ | ❌ | Native format | ❌ | ❌ | ❌ | ❌ |
-| **Cross-agent** | ✅ MCP + REST | API calls | Within runtime | Standalone | MCP + API | Standalone | Python API | Multi-agent shared |
-| **Audit trail** | ✅ All mutations logged | ❌ | Limited | ❌ | ❌ | ❌ | ❌ | ❌ |
-| **Language SDKs** | Any (REST + MCP) | Python + TS | Python only | API | Python + TS | Python | Python only | Node |
-
----
-
-## Token Efficiency
-
-The main reason to use persistent memory at all: token cost. Here's what one year of heavy agent use looks like across approaches.
-
-| Approach | Tokens / year | Cost / year | Notes |
-|---|---|---|---|
-| Paste full history into context | 19.5M+ | Impossible | Exceeds context window after ~200 observations |
-| LLM-summarized memory (extraction-based) | ~650K | ~$500 | Lossy — summarization drops detail |
-| **agentmemory (API embeddings)** | **~170K** | **~$10** | Token-budgeted, only relevant memories injected |
-| **agentmemory (local embeddings)** | **~170K** | **$0** | `all-MiniLM-L6-v2` runs in-process |
-| supermemory | Not published | Cloud pricing | Managed API, no local token budget |
-| Mem0 | Varies by integration | Varies | Extraction-based, no token budget |
-
-**agentmemory ships with a built-in token savings calculator.** Run `npx @agentmemory/agentmemory status` after a few sessions and you'll see exactly how many tokens you've saved vs. pasting the full history.
-
----
-
-## What Each Tool Is Best At
-
-This isn't a "agentmemory wins everything" page. Different tools solve different problems.
-
-**Choose agentmemory if you want:**
-- Automatic capture with zero manual `add()` calls
-- MCP server that works across Claude Code, Cursor, Codex, Gemini CLI, etc.
-- Hybrid BM25 + vector + graph search
-- Real-time viewer to see what your agent is learning
-- Self-hostable with zero external databases
-- Privacy filtering on API keys and secrets
-- Multi-agent coordination (leases, signals, routines)
-
-**Choose Mem0 if you want:**
-- Framework-agnostic API to bolt onto an existing agent
-- Managed cloud option with a dashboard
-- Python + TypeScript SDKs for direct integration
-- Entity/relationship extraction as the primary abstraction
-
-**Choose Letta/MemGPT if you want:**
-- A full agent runtime, not just memory
-- OS-inspired memory tiers (core/archival/recall)
-- Agents that self-edit their memory via function calls
-- Long-running conversational agents (weeks/months)
-
-**Choose Khoj if you want:**
-- A personal AI second brain, not agent infrastructure
-- Document-first search over your files and the web
-- Obsidian/Notion/Emacs integrations
-- Scheduled automations and research tasks
-
-**Choose supermemory if you want:**
-- A managed memory API with server-side auto-extraction and automatic forgetting
-- Drop-in wrappers for major AI frameworks (Vercel AI, LangChain, LangGraph)
-- A hosted dashboard with no infrastructure to run yourself
-- RAG plus memory served from a single query
-
-**Choose MemPalace if you want:**
-- A simple, free, open-source vector memory store
-- To chase its self-reported retrieval benchmark (we have not reproduced it)
-- Pure retrieval over agent workflow features
-- Note: no auto-capture, no MCP, no multi-agent coordination, so you wire all integration yourself
-
-**Choose oracleagentmemory if you want:**
-- You already run on Oracle AI Database and want memory inside it
-- Enterprise Oracle stack with vector search in the same database
-- LLM-backed extraction and are fine paying for a frontier model (their benchmark used GPT-5.5)
-- Note: Python-only, Oracle Database required, no MCP, no real-time viewer
-
-**Choose Hippo if you want:**
-- Biologically-inspired memory model (decay, consolidation, sleep)
-- Multi-agent shared memory as a primary feature
-- "Forget by default, earn persistence through use" philosophy
-
-**Choose TencentDB Agent Memory if you want:**
-- Team-level shared memory: conversations, docs, and code turned into four asset types (Chat Memory, Skill, Wiki, CodeGraph) with team roles and ownership
-- Zero-integration capture via an LLM proxy (point the agent's base URL at it; no hooks or MCP required)
-- CodeGraph impact analysis (symbols, call relationships) alongside memory
-- Note: the proxy sits in front of every model call, deployment is a multi-service Docker stack (Core + Hub + Proxy), the published benchmark is PersonaMem (76%, self-reported), and automated memory routing is still in progress per their README
-
-**Choose Zep / Graphiti if you want:**
-- A temporal knowledge graph: facts carry a time dimension, so "what was true when" is a first-class query
-- The strongest published temporal-query results (LongMemEval 63.8%)
-- Note: graph construction runs in the background, so freshly ingested facts can take time to become retrievable, and per-conversation memory footprint is reported to run far above extraction-based systems
-
-**Choose Cognee if you want:**
-- Knowledge-graph construction from documents and structured data before query time
-- Entity-relationship extraction as the primary product rather than session capture
-- Note: Python-only, and built for document ingestion rather than coding-agent memory
-
----
-
-## Running Your Own Benchmarks
-
-We encourage you to measure this yourself rather than trust any README. Here's how:
-
-```bash
-# Clone the repo
-git clone https://github.com/rohitg00/agentmemory.git
-cd agentmemory && npm install
-
-# Run LongMemEval-S
-npm run bench:longmemeval
-
-# Run quality benchmark (240 observations, 20 queries)
-npm run bench:quality
-
-# Run scale benchmark
-npm run bench:scale
-
-# Run real embeddings benchmark
-npm run bench:real-embeddings
-```
-
-Results land in `benchmark/results/`. All scripts, datasets, and results are committed for reproducibility.
-
----
-
-## Corrections Welcome
-
-If you maintain one of these tools and we got a number wrong, please open an issue or PR. We'd rather have accurate numbers than convenient ones.
-
-If you want to add your tool to this comparison, open a PR with:
-1. A link to your benchmark methodology
-2. The metric and dataset you're measuring on
-3. A commit hash / version so we can reproduce
-
-**Sources:**
-- Mem0 LoCoMo benchmark: [mem0.ai blog](https://mem0.ai)
-- Letta LoCoMo benchmark: [letta.com/blog/benchmarking-ai-agent-memory](https://letta.com/blog/benchmarking-ai-agent-memory)
-- LongMemEval paper: [arxiv.org/abs/2410.10813](https://arxiv.org/abs/2410.10813)
-- LoCoMo paper: [snap-stanford.github.io/LoCoMo](https://snap-stanford.github.io/LoCoMo/)
+For commercial services, use their current plans instead of a copied monthly price: [Mem0](https://mem0.ai/pricing), [Supermemory](https://supermemory.ai/pricing), [Zep](https://www.getzep.com/pricing/), [Letta](https://docs.letta.com/), [Cognee](https://docs.cognee.ai/cognee-cloud/overview), and [claude-mem](https://github.com/thedotmack/claude-mem). Check the chosen deployment's provider, telemetry, retention and access controls before using private data.
diff --git a/benchmark/LONGMEMEVAL.md b/benchmark/LONGMEMEVAL.md
index 9b1afa41b6..9e0c57483f 100644
--- a/benchmark/LONGMEMEVAL.md
+++ b/benchmark/LONGMEMEVAL.md
@@ -1,4 +1,6 @@
-# LongMemEval-S Benchmark Results
+# Historical LongMemEval-S Benchmark Results
+
+This report uses historical **any-hit@k** scoring and an older input pipeline. It is not a 0.9.31 strict-recall result. See the [current evaluation and receipts](../docs/benchmarks/2026-10-28-retrieval-evaluation.md); the metrics and denominators differ.
[LongMemEval](https://arxiv.org/abs/2410.10813) (ICLR 2025) is an academic benchmark for evaluating long-term memory in chat assistants. It tests 5 core abilities: information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention.
diff --git a/docs/benchmarks/2026-10-28-retrieval-evaluation.md b/docs/benchmarks/2026-10-28-retrieval-evaluation.md
new file mode 100644
index 0000000000..c1d4278df1
--- /dev/null
+++ b/docs/benchmarks/2026-10-28-retrieval-evaluation.md
@@ -0,0 +1,57 @@
+# Retrieval evaluation, 2026-10-28
+
+AgentMemory 0.9.31 development checkout, revision `3f8784d10c2e996873d4392b85226a5d042def0e`. These runs exercise the production BM25 `SearchIndex` directly and a lexical baseline over identical supplied sessions. They do not exercise the daemon, hooks, compression, vectors, graph, persistence, or an answer-generating model.
+
+## Results
+
+| Dataset | Adapter | Strict recall-all@5 | Fractional recall@5 | Any-hit@5 | Completed / attempted |
+|---|---|---:|---:|---:|---:|
+| LongMemEval-S | BM25 component | 84.7% (398/470) | 91.7% | 96.2% (452/470) | 500/500 |
+| LongMemEval-S | grep baseline | 72.1% (339/470) | 80.8% | 88.7% (417/470) | 500/500 |
+| coding-agent-life-v1 | BM25 component | 93.3% (14/15) | 93.3% | 93.3% (14/15) | 15/15 |
+| coding-agent-life-v1 | grep baseline | 100% (15/15) | 100% | 100% (15/15) | 15/15 |
+
+All 1,030 adapter attempts completed with zero execution failures. LongMemEval has 470 answerable questions and 30 unanswerable questions. The latter are included in coverage but excluded from retrieval metrics; this run does not measure abstention quality. The small coding fixture is a development regression set, and the simpler baseline does better on it.
+
+**Strict recall-all@5** requires every gold source session among the first five distinct returned sessions. **Fractional recall@5** averages the fraction of required sources found. **Any-hit@5** needs only one. Multi-source questions can therefore have an any-hit success while still missing necessary evidence.
+
+| LongMemEval question type | Answerable | BM25 strict successes | grep strict successes |
+|---|---:|---:|---:|
+| Single-session user | 64 | 63 | 60 |
+| Single-session assistant | 56 | 55 | 50 |
+| Single-session preference | 30 | 25 | 14 |
+| Knowledge update | 72 | 70 | 66 |
+| Temporal reasoning | 127 | 99 | 85 |
+| Multi-session | 121 | 86 | 64 |
+
+Multi-session retrieval remains the weakest BM25 category: 86/121 strict successes (71.1%). A complete answer-quality evaluation still needs a fixed reader, judge, model, prompt, context budget, and matched ingestion path.
+
+## Reproduce
+
+Run from the repository root with dependencies installed:
+
+```sh
+mkdir -p benchmark/data
+curl -fL https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/98d7416c24c778c2fee6e6f3006e7a073259d48f/longmemeval_s_cleaned.json -o benchmark/data/longmemeval_s_cleaned.json
+shasum -a 256 benchmark/data/longmemeval_s_cleaned.json
+npm run eval:longmemeval -- --data benchmark/data/longmemeval_s_cleaned.json --adapters grep,bm25
+npm run eval:coding-life -- --adapters grep,bm25
+```
+
+Expected dataset SHA256: `d6f21ea9d60a0d56f34a05b609c79c88a451d2ae03597821ea3d5a9678c3a442`. The external dataset is not redistributed. The loader removes answer labels, uses opaque source IDs, retains supplied dates, and coalesces 13 identical repeated distractor IDs. Each question gets a fresh index. See the [evaluation guide](../../eval/README.md) for the input transformation and metric contract.
+
+The run used Node v25.6.1 on macOS arm64. Manifests record source, installed dependency and dataset hashes. `dirty: true` records an untracked dependency symlink; tracked source was unchanged during the run. Diagnostic query timings exclude ingestion, initialization and cleanup, have no warmup or repeated trials, and are not end-to-end latency claims.
+
+## Receipts and scope
+
+- [LongMemEval summary](receipts/2026-10-28/longmemeval/summary.json) and [manifest](receipts/2026-10-28/longmemeval/manifest.json).
+- [LongMemEval per-question scores](receipts/2026-10-28/longmemeval/scores.ndjson.gz) and [returned-source evidence](receipts/2026-10-28/longmemeval/evidence.ndjson.gz).
+- [Coding summary](receipts/2026-10-28/coding/summary.json), [manifest](receipts/2026-10-28/coding/manifest.json), [scores](receipts/2026-10-28/coding/scores.ndjson.gz), and [evidence](receipts/2026-10-28/coding/evidence.ndjson.gz).
+
+Receipts contain public fixture identifiers, content hashes and scores, without transcript text or personal memory data. Decompress the NDJSON with `gzip -dc `.
+
+These results do not establish a speedup or accuracy gain over 0.9.30. The historical [hybrid LongMemEval report](../../benchmark/LONGMEMEVAL.md) used any-hit scoring and a different input pipeline; its 95.2% headline cannot be compared to strict recall-all here. The [v0.6.0 token experiment](../../benchmark/REAL-EMBEDDINGS.md) and [May coding benchmark](2026-05-20-coding-agent-life-v1.md) remain historical measurements, not 0.9.31 guarantees. No third-party memory product was run in this evaluation.
+
+## What changed in 0.9.31
+
+The [release notes](https://github.com/rohitg00/agentmemory/releases/tag/v0.9.31) document scoped saved-memory recall, incremental session summaries, paged exports, validated imports, native Pi/OMP/DeepSeek Harness integrations, Cursor workspace attribution, and compression/recovery fixes. These are implementation improvements. Lifecycle correctness and packaged startup/restart behavior are tested separately by the [evaluation suites](../../eval/README.md) and [release gate](../../scripts/release-gate/); this component run does not certify them.
diff --git a/docs/benchmarks/receipts/2026-10-28/coding/evidence.ndjson.gz b/docs/benchmarks/receipts/2026-10-28/coding/evidence.ndjson.gz
new file mode 100644
index 0000000000000000000000000000000000000000..a5ac31f989d5a995c5ed0979541dab04eb043c98
GIT binary patch
literal 3680
zcmV-m4xjNKiwFP!00002|IJ*>k{q`Yy!$iuOiiFr51{l1KI^DZIPjDmVUiL_9UKb%
zyAyOs+O_7jq&&G&*Jm^tO#q>o@?qjmXnH1V&a
zseznQ4&68ujZ--&zcaWRB`FHJiOo&4^PBk!Qa+nu%zS
zj*w!hrL-8$utdYJj>cl@SaRl(M=o>D6vbklBehamVqq*)@qGC)=)@&N=fN1!T_D1N}y76YGRIQNXT7}fvB4SP%W1c
zz1#3QqV=3n%Y;$yGUlk}8b(Zj|H9}|gs~FTQlkPC6bhN$LJV>UVfcB3(b%
z_k15=E4=8Mq4!C$iw|ET%w}PrT7hE8AyqK|Juwn?a=sph
zuNFb)!=TR#LzTHwKGqR4*U>u)0*bBQ6@~>DattlyNf4r=IVw2WpA(lo)!&3hP}KHp
zlaaP!EkH<^SM%0ppt!iOhh))aJ>*;}Wu%AREcvV`5#yjN3Ydbzm;nSz?!80R7?-fI
z2s%#~dnRE<>?IYIrUVtqOXNn5bYG#DY$63t08?&`@D9}HdfhH<7KV#cxvJ9#gJcDR
z$H+ZMqLQ0VXF(Bh&!N?x8u$m*QKCyq7_Af(oSBQ(%UZP3aInjH(5RZ1-L5OT$CTF5
zF`Tt5DVA<{CZqrsmyNaPCHmHgikvW+5lR?}ICm<RL+a&1pGYBw
z-3>LW3-q4^$wF6U$JdA9(r8dThpIO+b<9^?8hke$lhZ9_S78CgONWRV$g&$cvw+|@&7$KMUr%V
zdvGhlxWK(RSX*mmlL;QMvRtosir=A7W=mpB@S=`-jRAFn$Yg_WZ}^M&^b7*}aF2j+
znWYt-tF&JlP0zK!V`DB{lO=X2>0D(TvlG;3!}B9*rl|}V%<92wJcKa*JixSo=x
zWhjCx%DBV`Si#A_*y1@QsPQ~A@eINghc0*IJ|+8!8T59=Lz%N^W$vrwGjpt=;(zEz-Jqu6XEZ|OCbYL#npgv
zZ>XVOsir$s+ogvO*+`3UM~wA@dIupl#N^s!oS+g$SyrV=Setb
zUw}EXK!{E;nikdz0>jcBvkxvh-lru#GdvTm&m16_(IdkZXHQo=Ic`L$x77IRdrnoQPoSoIZQqdg+Lno_WM<0giQ8
z_A0Sfq{NwFNSW^PU1|>z
zf?)~1HZe280t2kaSiq2FGHM(?Kzj3O)H0RQkEt*hL3adUo_WL!JQjqg#*Bg;3T5Ih
z)8Zr5rZgx8(Y?|Xk`MFa>U3xSg3=J*j@O)+v(*;yz}jN`~Jg!zNv4c
ze$(E+{ptNX{P4&->qIAwQ3H}b=dir~o`;1uuZ_zoX3Rk&QLuGzvm?)LU$K`Ht;!JK
zX)GWs%Z@qej!W){EJ7MQWH5qR@}7mi7z_q?;ufH#=dc8^*B;+`KS>7j@WB>fL-3OQ
z?k>caVCKe`$ct2(~*`ADU#Z}?m
zLKLj-ewSAHC1|}vsXmd4AR`0%hIrW`+>RTu84@DFL#ByxIAccjb>xTngqN@=$a*nh9bi27Yfsac
z;N`~6AL(R>1QrH@CYZ792>jda7kUC(Xr*#bP|e9+RXYaM23&7IvSRqBcn+=Xs|tkmh|$>htmfU3Zxh%I9W}w|9>d@i+vALjXo4chXyWkhrdv3ud%8)!l$T%A
zV{*HzLrz!{WP5_FB<8crvsIVJ?G6|ySr}^IOCTj9d*D4=ViL5q0b-p%NzaPU13>E&
zmjoOW9DHV!&75sVdwxfc%#lL-jy9D8_kPk#P;=vM7C$+d>1)pnt`Nc5vAgAto{1yD
zb-VfIP#0$r8n#1fQ;m=w1%@P6l+Gu^L)C%~!YC!e(Wf01#~tlC$h%7rF)Ai>JQ44A
zV3J>gnj81c*a=ASux>)a5ZSrq+3dvlcIOJ8^4d`}W?jf<%R62@xZUOhNyZQl(IFW^
zvb^UN@my#OrwmO4)y=>tas-$7N_LGWto4}725@OG>T$xaOn(*ovi=gh+_*glC377^
z!{sT|qMqAc+3v2^&6bpS0_r(s02|5=ywg3uJ7paY+9D9LlGoN!=hfHsq@u%oMx+u&
zVGYK4&u<^P-RyKyT^d3STH8@HUsvngp>)t|%Ljh5J5yJs28hyrG5&+qZM+0>$I`%!^R
zz!-`)?Ro9(c5{@R5>}3ZCSP75+UshTI#BE4otmk1cu>mt#nLD87v;To*Y4?G`qE1X
z+jcb(o#0*o0!KyFb60~7Y;uwlr-fz4+H;QQwB(UZPp2G^%CK(93B~u^Y2AhmFE{&+
zP6>0sKVjX@U*Fj79QJmL)@iK+R9pb^VvtR4%iq%^^@Xhbs@AEgqZ`WOnCl+NI_s>^
zx%}k7Y3=F-&a4qMOMCVn-qE0QN=Up0tSC}P%DyAC?r1w(w;fY(hRy}2eAjQ3xY=q2
ziwC+wM27?~kr1pszu@U+mmG!%h8V{fZEb#}wA&2?dXbw!$G2#qqONNTq7iutW8nCz
zdbR{|P1ry)T8dZpyxVlUYl=^6CmP60nFx~nT;|*G%cWW<1IPwsdxs%i6NGenGCX9n
z456Py@Si=Kb?#_@+e3-p)AaR)+}*kr>~`}NpAcFRI9WsN`P?SShF`97t5I!TWPmpe
z>#_>mh39g8XN9>76>bp4QzEMGlW*!hVaA+|80Br9
yx$U`a-cxCg6|Cx3268(pN77mz5AU;lGg2_u^EpZN^?Qe)?tcM`-3uT3WB>qc10*Q`
literal 0
HcmV?d00001
diff --git a/docs/benchmarks/receipts/2026-10-28/coding/manifest.json b/docs/benchmarks/receipts/2026-10-28/coding/manifest.json
new file mode 100644
index 0000000000..423dcc7158
--- /dev/null
+++ b/docs/benchmarks/receipts/2026-10-28/coding/manifest.json
@@ -0,0 +1,80 @@
+{
+ "schemaVersion": 1,
+ "metricVersion": "distinct-session-retrieval-v2",
+ "inputTransform": "role-flattened; opaque source IDs; all supplied dates retained; identical repeated IDs coalesced (v1)",
+ "benchmark": "coding-agent-life-v1",
+ "scope": "supplied-session retrieval; no reader, judge, or capture-quality score",
+ "startedAt": "2026-10-28T07:28:33.253Z",
+ "source": {
+ "revision": "3f8784d10c2e996873d4392b85226a5d042def0e",
+ "dirty": true,
+ "contentSha256": "25c0fc4f610b612675600ebac66da592716081b7cf97b34c018b417f1b34520f",
+ "dependencies": {
+ "@anthropic-ai/claude-agent-sdk": "0.3.308",
+ "@anthropic-ai/sdk": "0.100.1",
+ "@clack/prompts": "1.8.1",
+ "@huggingface/transformers": "4.3.1",
+ "@iii-dev/helpers": "0.25.0",
+ "@node-rs/jieba": "2.0.4",
+ "@types/node": "26.6.5",
+ "dotenv": "18.0.8",
+ "iii-sdk": "0.25.0",
+ "json5": "2.2.3",
+ "picocolors": "1.1.1",
+ "sigstore": null,
+ "tiny-segmenter": "0.2.0",
+ "tsdown": "0.21.10",
+ "tsx": "4.23.15",
+ "typescript": "6.0.3",
+ "vitest": "4.1.11",
+ "ws": "8.22.0",
+ "yaml": "2.9.1",
+ "zod": "4.6.7"
+ },
+ "installedLockSha256": "d8a23fe5cc06b1ae740ed1eb72594e9af9f2d52231a7431fa2d509d54e9f0950",
+ "builtArtifactSha256": null,
+ "engine": null
+ },
+ "datasetFiles": [
+ {
+ "name": "sessions.json",
+ "sha256": "d32f0c3c35a35e6c5aa0ac1fd0bd92e1f413c132efe3b1262038595589df3f16"
+ },
+ {
+ "name": "queries.json",
+ "sha256": "4f36fb00ea1e0ce3ccbefea9348fc4da027201a26a1fd7b5349391bcc69a24b4"
+ }
+ ],
+ "datasetSha256": "7c37494bd6c3211e937c132b7501a594151f56337f933d8c617f79d4bf4c0701",
+ "selectedQuestionIds": [
+ "q-001",
+ "q-002",
+ "q-003",
+ "q-004",
+ "q-005",
+ "q-006",
+ "q-007",
+ "q-008",
+ "q-009",
+ "q-010",
+ "q-011",
+ "q-012",
+ "q-013",
+ "q-014",
+ "q-015"
+ ],
+ "coalescedRepeatedSources": 0,
+ "adapters": [
+ "grep",
+ "agentmemory-bm25-component"
+ ],
+ "k": 5,
+ "retrievalUnit": "top k distinct source sessions",
+ "config": {},
+ "runtime": {
+ "node": "v25.6.1",
+ "platform": "darwin",
+ "arch": "arm64"
+ },
+ "planned": 30
+}
diff --git a/docs/benchmarks/receipts/2026-10-28/coding/scores.ndjson.gz b/docs/benchmarks/receipts/2026-10-28/coding/scores.ndjson.gz
new file mode 100644
index 0000000000000000000000000000000000000000..45a4dcc3f23244c78770acf86ca2e39301960577
GIT binary patch
literal 988
zcmV<210(z&iwFP!00002|J|9-Zrex<#_#(WI>*6q{!ea24@FVzWw9@y*yB`)EV?&JPdm@k?9p
z{wkZt^WnM-zS%zi*`J;bPB1Ij+XWKO<4e819J%Tspy(Qvd$(lGEXnbgE|Gm5b*i
z-tl3=NYEo6BW2D9loi~)URFy<@)47KI=?Tsq3u?eLX2CkQU)<+az@NBJ2KMB8B0Qv
zlaTfv-?HRtkc`wMqkuP8SZBb}r5q4-#&yiIbDEKS
z^dz67@Aix5!?OfJl^Io7NN_~{XzpN4F`*=oe9R=D@>s1;g9Jx~8Iwa8t_eQerus76
zGFqwaB%s~>3i({54_1*=l>VeX56F|wANQLy`dp!vNNwB!CD2%dG6G_p35pP~pJ?^9
zFMGejEMgAS8zaXwSQ3(qQf{mxJ3SD?^kJw&6%lRAp)c#J5R%sAQue~tLCC|)sf1+VZpIbld==R
zVPnVEM$3+zDyh!@Q~9SAmK3&jHx>*BMx_p<){c`QH#8z*b`JhiR*qn51VDHwz?{SO#>4~)G5+n04eO{^K}`n^RYg>ew;PsxlLGEJ)os{3pX$fA=Ck)gJa1uC+Y
zA8v)P-0_jDoXZngG&S7fE;c!(R7OpMO4NYaS!|L;G0V&WoFHPR@!TAMTR52~qpsOm
zjFV+&nYlCrnA90Dj=PH*#||mx@KP3Dmyot!&P-mHC(j^xoKDZ+t}I*s1xtgw6$nA;9PvEOq`@I>#A^6l9?Y_H4J3~<8Uan5U59ZK06mW>^I>H2
zs^gjBmwRQNXx5lGXD~3=v}qPzlPZ0^b4KLN63*X&EvDo<}F;F-I^1Wo)~Y!(Ga|-_j2Q8+fBf(Nv;X{;|M#E&>A$W&{_&3;%h+@D
zKmOs5zWTrXW&er4{9k|C|H5DVuYdmk)_>c7{`2#*fBxei|N9?*{`GJF)1RJCUo)=g
z{2al#6GG4ZkN@!xfBb3x<^TNYpEh6q-~aXBKk)pjfB*14BlaBKYH=nTF(h~P=#$_0
z=lRcn+JF3G{D(j0KmXHT_CNjQzx_Ymr_6^#cG*r!`?TXysPn8G&Z%vje&!f6lLths{i80@s{dJFxJfBT2ePxJaJx2IWWh7z`KVOg#|M74CFdCZq;CA$@H6V;VI
zCa1(NzHf(itdBmh;k{OIb)CFIbbCa5eVSNSIu)0YN>0&INg>AI&h=?9U0+V9w`?5S9{b4?qdM{U+IjUDVok44!}?(&x#N?AT?vWPfe~ir$^kr_>#_F7u>T)YS5q=gn
zWY{6j_4wfZ(I5Foo9{Lga<5pfCv4#q@@Dr6zIt7KX`CGDvwU)6#OnOmzFOYF?e0(J
zpINU~497WkDY(~uu*S_V7-z5dQ|gurX(@njd{o4I^gi`B3-e_hML!-BwOUF{
zGtatKtFs^d^D$@fw%yq87R-syyyIyTz~Kpzm|5e+?oVTL27Y;o
zJ6CIG5bS2bdzCUfTjZu|X1J8L8^_4EE3FpV