A deep research agent kernel with long-term memory. Every claim is weighed on the scales of evidence before it is reported. The kernel is middleware-first: memory, skills, sandbox, context governance, and reflection are pluggable cross-cutting concerns layered around a ReAct loop — not embedded agent logic. Each layer is independently designed, independently tested, independently verifiable, guarded by 2,700+ tests.
| A real terminal interface | Full-screen TUI (Textual) with side panel, live token usage in the status bar, slash-command palette, and settings screen — plus a traditional scrolling CLI (prompt_toolkit + rich) and a one-shot anubis run mode. |
| Five-layer long-term memory | Cognitive-science-inspired memory: Ebbinghaus decay with lazy strength computation (no background tasks), BM25 hybrid retrieval with write-back reinforcement, per-call HumanMessage recall injection that protects prompt caching, and non-blocking auto-consolidation that logs-and-skips instead of blocking the loop. |
| Context engineering governance | A live token budget resolved against the real model window at call time — no hardcoded thresholds. Dual strategies (compaction + externalization) keep long research sessions from overflowing without losing critical information. |
| Delegates and parallelizes | Route sub-tasks to external specialists (pi / codex / claude) via MCP, or spawn self-copy subagents with isolated context that share the parent's Docker sandbox — same artifacts, zero re-acquisition. |
| Sandbox isolation with path enforcement | Docker and local providers. Write-path whitelisting forces agent output into the mount area — nothing is lost to --rm. Warm pool, idle auto-destroy, cross-process lock, WSL2 executor. |
| A closed skill-evolution loop | Skills are prompt-level research knowledge bundles (not executable functions). When the effective rate drops below threshold, LLM-driven mutation runs through score gating and GitRatchet rollback — evolution can only move forward. |
| Use any model, degrade gracefully | Role-based routing chain (researcher / reporter) with automatic fallback on rate limits, timeouts, and 5xx — DeepSeek always sits at the chain tail as the ultimate fallback. |
| Evaluable by design | 2,700+ tests guard every layer independently. A structured run journal records skill selections, memory encodes, compactions, and token budgets for post-hoc analysis. |
flowchart LR
U["Question"] --> P[prepare] --> L["LeaderAgent<br/>ReAct Loop"] --> F[finalize] --> A["Answer + Evidence"]
subgraph MW["22 Middleware · every lifecycle hook"]
direction TB
BM[before_model<br/>memory recall · skill injection] --> M[model call]
M --> AM[after_model<br/>consolidation · context governance]
end
L <--> MW
MW --> T[tool calls]
T --> SB["Docker Sandbox<br/>path-guarded"]
T --> MC["MCP Servers<br/>stdio / sse / http"]
T --> BT["Builtin Tools<br/>ddg · files · skills"]
flowchart TB
L5["L5 · Auto-Consolidation<br/>daemon worker · queue · log-and-skip"] -->|"every N turns"| L4
L4["L4 · Middleware + Bootstrap<br/>per-call HumanMessage · set_turn_id"] -->|"recall"| L3
L3["L3 · Store + Retriever<br/>MarkdownFileStore · BM25 · write-back reinforce"] -->|"strategies"| L2
L2["L2 · Default Strategies<br/>Ebbinghaus decay · lazy strength · forgotten filter"] -->|"schema"| L1
L1["L1 · Schema + Protocol<br/>MemoryTrace · 5 atomic ops · tools have no LLM"]
Memory strength is computed lazily at retrieve time — no background tasks, no scheduler to babysit:
strength = base × (1 − decay)^hours + log(1 + access) × 0.1 + importance × 0.05
Requirements: Python 3.12+, optional Docker for the sandbox provider.
git clone https://github.com/fxl112233/anubis-agent.git
cd anubis-agent
pip install -e .
# First run (or anytime .env is missing) launches a setup wizard.
# Alternatively, copy the template and fill in at least one provider key:
cp .env.example .env
# Full-screen TUI (default)
anubis
# Traditional scrolling CLI
anubis cli
# One-shot deep research, then exit
anubis run "Compare LK-99 replication attempts since 2024" --no-expert| Command | Description |
|---|---|
anubis |
Full-screen TUI (Textual): conversation + side panel, live token status bar, / command palette |
anubis cli |
Scrolling CLI with rich streaming and slash-command completion |
anubis run "<question>" |
One-shot research run; prints the final report and run ID. --expert (default) enables deep-research mode, --no-expert is lightweight |
anubis run ... --thread-id <id> |
Continue an existing research thread |
anubis run ... --no-artifact |
Skip writing the final report file |
| Command | Description |
|---|---|
/expert / /default |
Switch between deep-research and lightweight modes (thread state is kept) |
/model <provider> [model] |
Hot-swap the LLM mid-session |
/report |
Manually trigger report synthesis |
/skill <name> |
Override the active skill set |
/mcp reload |
Reload MCP tools without restarting |
/help, /exit |
Show all commands / quit |
All configuration lives in .env (template: .env.example). At least one provider key is required; each provider can be disabled with {NAME}_ENABLED=false and its default model overridden with {NAME}_MODEL.
| Provider | Env var | Notes |
|---|---|---|
| DeepSeek | DEEPSEEK_API_KEY |
Default provider; tail of every fallback chain |
| OpenAI | OPENAI_API_KEY |
Any OpenAI-compatible base URL supported |
| Qwen | QWEN_API_KEY |
OpenAI-compatible endpoint |
| Anthropic Claude | ANTHROPIC_API_KEY |
Requires pip install langchain-anthropic |
| Google Gemini | GEMINI_API_KEY |
Requires pip install langchain-google-genai |
| Moonshot / Kimi | MOONSHOT_API_KEY |
OpenAI-compatible endpoint |
Model behavior profiles (routing, budgets, expert mode) live in anubis/backend/agents/config/profiles/: fast.yaml, general.yaml, expert.yaml.
anubis/backend/
├── app/ # CLI, TUI (Textual), setup wizard, stream services
├── agents/
│ ├── leader/ # LeaderAgent ReAct loop + prompts
│ ├── middlewares/ # 22 lifecycle middlewares (memory, skills, governance…)
│ ├── memory/ # five-layer long-term memory stack
│ ├── context_engineering/ # live token budget, compaction + externalization
│ ├── skill/ # builtin skills + closed evolution loop
│ ├── multiagent/ # MCP delegation, self-copy subagents, specialist runtimes
│ ├── sandbox/ # Docker/local sandboxes with path enforcement
│ ├── mcp/ # MCP client, loader, health, credential guards
│ ├── config/ # provider config, model router, fallback chain
│ ├── reporting/ # report synthesis
│ ├── journal/ # structured run journal
│ └── observability/ # events and traces
└── tests/ # 2,700+ tests across every layer
# Full test suite
pytest
# Run a single layer (unit tests are organized per module)
pytest anubis/backend/tests/v1/unit/memory
# Integration tests (CLI chat/run, agent loop e2e)
pytest anubis/backend/tests/v1/integrationMIT License — 随意使用、修改、商用，保留 License 声明即可。
如果这个项目对你有帮助，欢迎点个 ⭐ Star！
有问题欢迎提 Issue