Skip to content
This repository was archived by the owner on Nov 16, 2026. It is now read-only.
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .cursor-plugin/plugin.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "agentmemory",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"description": "Persistent memory for Cursor: auto-captures sessions, prompts, and tool use via hooks, recalls with hybrid BM25 + vector + graph search, and exposes 55 memory tools over MCP. Keyless by default, runs entirely on your machine.",
"author": {
"name": "Rohit Ghumare",
Expand Down
12 changes: 12 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -1,3 +1,11 @@
## 0.9.33-daniel.75 — Additive memory-pressure recovery

- Preserve the existing available-memory normal/recovery paths; Linux PSI adds a separately qualified recovery path rather than vetoing previously allowed work.
- Keep independent legacy and PSI observation windows, invalidate interrupted legacy evidence at every activation read, and reject malformed PSI fields instead of treating them as zero.
- Preserve critical protection, recovery-attempt history, other pause holds, and final durability/activation checks. Existing application data and capture originals are not rewritten by this policy change.
- Correct the `.74` candidate's additional PSI veto. The `.75` package passed isolated tests and installed-package save/search/normal-shutdown verification, and was subsequently deployed. Live observation does not guarantee safety against future allocation bursts.
- Policy and limitations: `docs/customizations/MEMORY-PRESSURE-BUDGET.md`.

## 0.9.32-daniel.68 — Sharded graph retrieval index

- Store new graph retrieval builds in per-build hash shards to stay below the native file namespace size limit.
Expand All @@ -22,6 +30,8 @@

## [Unreleased]

## [0.9.33-daniel.75] — 2026-11-14
## [0.9.33-daniel.74] — 2026-11-14
## [0.9.33-daniel.73] — 2026-11-14
- Stage existing SDK capture queue reference replacements in a private, quota-counted queue workspace, avoiding permanent ext4 directory growth from random temporary names in large queues. Preserve the existing 5 MiB admission and 8 KiB transition budgets, crash files, source originals, and conservative refusal when legacy bootstrap cannot fit. Source fix only; installed services are unchanged. See `docs/customizations/SDK-SPOOL-TRANSITION-WORKSPACE.md`.
- Make Ubuntu CI subprocess fixtures use the active Node executable. Validate finite capture deadlines under contention without weakening exact-once ACK/original preservation checks, and wait for real browser fixture completion instead of a fixed virtual-time DOM dump. Test-only changes; no operating service changes.
Expand Down Expand Up @@ -690,6 +700,8 @@ Wave release closing several breaking regressions reported against v0.9.26, plus
- `/agentmemory:forget` skill still calls `memory_governance_delete` which only touches `KV.memories` and never observations ([#833](https://github.com/rohitg00/agentmemory/issues/833)). Skill rewrite + new `memory_forget` MCP tool tracked separately.
- `crypto.randomUUID()` global-only on Node <19 ([#715](https://github.com/rohitg00/agentmemory/issues/715)). Drop-in import fix tracked.

[0.9.33-daniel.75]: https://github.com/rohitg00/agentmemory/compare/v0.9.33-daniel.74...v0.9.33-daniel.75
[0.9.33-daniel.74]: https://github.com/rohitg00/agentmemory/compare/v0.9.33-daniel.73...v0.9.33-daniel.74
[0.9.33-daniel.73]: https://github.com/rohitg00/agentmemory/compare/v0.9.33-daniel.72...v0.9.33-daniel.73
[0.9.33-daniel.72]: https://github.com/rohitg00/agentmemory/compare/v0.9.33-daniel.71...v0.9.33-daniel.72
[0.9.33-daniel.71]: https://github.com/rohitg00/agentmemory/compare/v0.9.33-daniel.70...v0.9.33-daniel.71
Expand Down
73 changes: 73 additions & 0 deletions docs/customizations/MEMORY-PRESSURE-BUDGET.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,73 @@
# Additive memory-pressure recovery

Status: released as `0.9.33-daniel.75` after isolated tests and installed-package
verification. Existing data and capture originals are not migrated by this policy.
No host pressure experiment was performed. Earlier release artifacts are unchanged.

## Why

Available-memory percentages alone can pause work even when measured stall pressure
is low. The additional path permits bounded nonzero stalls without claiming that
past host observations predict future concurrent-pipeline allocation bursts.

## Policy

Retain the original critical threshold `max(0.5 GiB, total * 5%)` and all critical
API refusal behavior. With valid Linux PSI only, permit an additional recovery
path after 120 seconds of fresh samples above
`max(3 GiB, critical + 2 GiB)`. Strictly greater is required at the floor.

The proposed operational stall budget is some <=1%, full <=0.1%, about 1400/146
times the observed whole-window ratios. These are candidate budgets, not proven
universal safety limits. Use counter deltas / actual monotonic elapsed time, plus
current avg10 and avg60 within the same budgets. Microstalls do not reset a window.
A large/sustained pressure reading or floor violation resets only the added
pressure window. The original policy remains authoritative: available >=18%
(or its existing reserve floor) is normal regardless of PSI, and available >22%
(or its existing recovery floor) permits recovery after 120 fresh seconds even
when PSI is high. PSI is an additional recovery opportunity below the original
reserve, never a veto on either legacy path. Once qualified, healthy pressure
prevents repeated pauses solely at 18%.

The .74 deployed candidate incorrectly added such a veto. A read-only trace
showed 10.51–11.00 decimal GB available with full stall 0.4615% over 120 seconds,
blocking legacy recovery. The .75 candidate removes the veto without increasing
the PSI budgets or bypassing critical, freshness, attempts or storage protection.
The original recovery timer and the PSI window are independent: a 119-second
quiet interval below the old recovery threshold plus one second above it with
high PSI cannot combine into an approved 120-second recovery. Either complete
path must independently qualify again during activation.

Missing/malformed PSI, unsafe or regressing counters, non-monotonic elapsed time,
and sampling gaps over 10 seconds disable the added path; use the original 18/22%
policy and rebuild the pressure window. Swap usage is not a zero-swap gate.
Sample counters are validated individually: full delta can microscopically exceed
some delta because kernel reads are not a simultaneous snapshot.

Existing wall-clock freshness checks, 120-second stable interval, three attempts,
backoff, activation-time rechecks, persisted manual/deployment/storage holds,
durability fences, and failed-activation restoration of the paused control state
are retained. This is not automatic software rollback: operational code rollback
requires separate approval, and data is never rewound. Recovery does not clear
other holds.

## Limits and replacement

A 5-second sampling cadence cannot prevent an allocation burst between samples.
No PSI threshold guarantees safe simultaneous pipeline startup. Final activation
checks reject pressure/floor changes before reopening admission, but post-start
burst behavior needs separately authorized operational observation. Upstream can
replace this customization only after equivalent fresh-window, critical protection,
hold preservation, and failed-activation control-state restoration tests pass
without losing data or retry history.

## Verification

Deterministic synthetic samples only, with injected monotonic time; no real memory
pressure induced. The focused budget tests cover boundaries, microstalls, missing
PSI, rollover, stale gaps, clock regression, current spikes, actual elapsed time,
compound holds, and post-fence pressure return. Existing protection/durability,
boot, build-profile, and fallback artifact suites are included. The full repository suite and exact-package start/save/search/normal-shutdown
checks also passed before deployment. Actual deployment retained existing capture
records; detailed operational evidence stays outside the public source tree.
Ongoing workload behavior and future allocation bursts still require observation.
2 changes: 1 addition & 1 deletion integrations/hermes/plugin.yaml
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
name: agentmemory
version: 0.9.33-daniel.73
version: 0.9.33-daniel.75
description: "Persistent memory for Hermes Agent with profile-scoped capture, recall, and native memory tools."
author: "Rohit Ghumare"
homepage: "https://github.com/rohitg00/agentmemory"
Expand Down
2 changes: 1 addition & 1 deletion integrations/omp/package.json
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@
"name": "agentmemory-omp-extension",
"private": true,
"type": "module",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"omp": {
"name": "agentmemory",
"description": "Persistent cross-session memory for omp \u2014 captures sessions, tools, approvals, and compaction; injects smart-search recall.",
Expand Down
2 changes: 1 addition & 1 deletion integrations/openclaw/openclaw.plugin.json
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@
"kind": "memory",
"name": "agentmemory",
"description": "Persistent cross-session memory for OpenClaw via agentmemory.",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"configSchema": {
"type": "object",
"additionalProperties": false,
Expand Down
2 changes: 1 addition & 1 deletion integrations/openclaw/package.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "agentmemory",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"type": "module",
"openclaw": {
"extensions": [
Expand Down
2 changes: 1 addition & 1 deletion integrations/openclaw/plugin.yaml
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
name: agentmemory
version: 0.9.33-daniel.73
version: 0.9.33-daniel.75
description: "Persistent cross-session memory for OpenClaw via agentmemory."
author: "Rohit Ghumare"
homepage: "https://github.com/rohitg00/agentmemory"
Expand Down
2 changes: 1 addition & 1 deletion package.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "@agentmemory/agentmemory",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"description": "Persistent memory for AI coding agents, powered by iii-engine's three primitives",
"type": "module",
"main": "dist/index.mjs",
Expand Down
4 changes: 2 additions & 2 deletions packages/mcp/package.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "@agentmemory/mcp",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"description": "Standalone MCP server for agentmemory — thin shim that re-exposes @agentmemory/agentmemory's MCP entrypoint",
"type": "module",
"bin": {
Expand Down Expand Up @@ -28,7 +28,7 @@
"homepage": "https://github.com/rohitg00/agentmemory#readme",
"bugs": "https://github.com/rohitg00/agentmemory/issues",
"dependencies": {
"@agentmemory/agentmemory": "0.9.33-daniel.73"
"@agentmemory/agentmemory": "0.9.33-daniel.75"
},
"publishConfig": {
"access": "public",
Expand Down
2 changes: 1 addition & 1 deletion plugin/.claude-plugin/plugin.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "agentmemory",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"description": "Persistent memory for AI coding agents -- captures tool usage, compresses via LLM, injects context into future sessions. 12 hooks, 55 MCP tools, 17 skills, real-time viewer.",
"author": {
"name": "Rohit Ghumare",
Expand Down
2 changes: 1 addition & 1 deletion plugin/.codex-plugin/plugin.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "agentmemory",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"description": "Local persistent memory for coding agents. Recall decisions, resume sessions, learn from corrections, and manage memory through 55 MCP tools and 17 skills. Optional capture through 6 Codex hooks.",
"author": {
"name": "Rohit Ghumare",
Expand Down
4 changes: 2 additions & 2 deletions plugin/.devin-plugin/plugin.json
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
{
"name": "agentmemory",
"displayName": "agentmemory",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"description": "Persistent memory for Devin CLI and Devin Desktop. Captures every session automatically, recalls with hybrid BM25 + vector + graph search, and exposes 55 memory tools over MCP. Keyless by default, runs on your machine. Not supported in cloud sessions.",
"logo": "assets/icon.svg",
"author": {
Expand All @@ -23,7 +23,7 @@
"command": "npx",
"args": [
"-y",
"@agentmemory/mcp@0.9.33-daniel.73"
"@agentmemory/mcp@0.9.33-daniel.75"
],
"env": {
"AGENTMEMORY_URL": "${AGENTMEMORY_URL:-http://localhost:3111}"
Expand Down
2 changes: 1 addition & 1 deletion plugin/opencode/plugin.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "agentmemory-capture",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"description": "OpenCode plugin for agentmemory — full Claude Code hook parity: session lifecycle (create/idle/status/compacted/update/diff/delete/error), messages & prompts (chat.message, message.updated user+assistant, message.removed), tool lifecycle (ToolPart states with timing), part tracking (subtask, step-finish, reasoning, file, patch, compaction, agent, retry), file enrichment pipeline, permissions, task tracking (w/ priority), commands, config & model tracking. 22 hooks, 2 slash commands.",
"author": {
"name": "Rohit Ghumare",
Expand Down
2 changes: 1 addition & 1 deletion plugin/plugin.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "agentmemory",
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"description": "Persistent memory for AI coding agents -- captures tool usage, compresses via LLM, injects context into future sessions. 12 hooks, 55 MCP tools, 17 skills, real-time viewer.",
"author": {
"name": "Rohit Ghumare",
Expand Down
2 changes: 1 addition & 1 deletion plugin/scripts/plugin-bridge.mjs
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ import { readFileSync, realpathSync } from "node:fs";
import { homedir } from "node:os";
import { join } from "node:path";
//#region src/version.ts
const VERSION = "0.9.33-daniel.73";
const VERSION = "0.9.33-daniel.75";
//#endregion
//#region src/secret-store.ts
const SECRET_KEY = "AGENTMEMORY_SECRET";
Expand Down
2 changes: 1 addition & 1 deletion scripts/docs-sync.state.json
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
{
"version": "0.9.33-daniel.73",
"version": "0.9.33-daniel.75",
"facts": {
"mcpTools": [
"55"
Expand Down
15 changes: 11 additions & 4 deletions scripts/personal-release/verify-recovery-artifact.mjs
Original file line number Diff line number Diff line change
Expand Up @@ -6,29 +6,36 @@ import {resolve} from 'node:path';
import {fileURLToPath} from 'node:url';
import {createHash} from 'node:crypto';
const ensure=(condition,code)=>{if(!condition)throw Error(code);};
const selectedNames=new Set(['GiB','SAFE_APIS','MEMORY_RECOVERY_ENABLED']);
const selectedNames=new Set(['GiB','SAFE_APIS','MEMORY_RECOVERY_ENABLED','PRESSURE_BUDGET']);
function isolatedGuard(bundle){
const file=ts.createSourceFile('reviewed-recovery-bundle.mjs',bundle,ts.ScriptTarget.Latest,true,ts.ScriptKind.JS);
ensure(file.parseDiagnostics.length===0,'compiled_guard_shape_unsupported');
const functions=file.statements.filter(s=>ts.isFunctionDeclaration(s)&&s.name?.text==='createMemoryPressureGuard');
ensure(functions.length===1&&functions[0].body,'compiled_guard_shape_unsupported');
const helpers=['validPressure','pressureWindow'].flatMap(name=>{
const matches=file.statements.filter(s=>ts.isFunctionDeclaration(s)&&s.name?.text===name);
ensure(matches.length<=1,'compiled_guard_shape_unsupported');
return matches.map(s=>s.getText(file).replace(/^export\s+/,''));
});
const constants=new Map();
for(const statement of file.statements){
if(!ts.isVariableStatement(statement))continue;
for(const declaration of statement.declarationList.declarations){
if(!ts.isIdentifier(declaration.name)||!selectedNames.has(declaration.name.text))continue;
const name=declaration.name.text,value=declaration.initializer;
const name=declaration.name.text,initializer=declaration.initializer;
const value=name==='PRESSURE_BUDGET'&&initializer&&ts.isCallExpression(initializer)&&initializer.expression.getText(file)==='Object.freeze'&&initializer.arguments.length===1?initializer.arguments[0]:initializer;
ensure(!constants.has(name)&&value,'compiled_guard_shape_unsupported');
let valid=false;
if(name==='MEMORY_RECOVERY_ENABLED')valid=value.kind===ts.SyntaxKind.TrueKeyword||value.kind===ts.SyntaxKind.FalseKeyword;
if(name==='PRESSURE_BUDGET')valid=ts.isObjectLiteralExpression(value)&&value.properties.length===2&&value.properties.every(p=>ts.isPropertyAssignment(p)&&ts.isIdentifier(p.name)&&ts.isNumericLiteral(p.initializer)&&((p.name.text==='some'&&Number(p.initializer.text)===1)||(p.name.text==='full'&&Number(p.initializer.text)===0.1)))&&new Set(value.properties.map(p=>p.name.text)).size===2;
if(name==='GiB')valid=/^1024\s*\*\*\s*3$/.test(value.getText(file))||ts.isNumericLiteral(value)&&Number(value.text)===1024**3;
if(name==='SAFE_APIS')valid=ts.isNewExpression(value)&&ts.isIdentifier(value.expression)&&value.expression.text==='Set'&&value.arguments?.length===1&&ts.isArrayLiteralExpression(value.arguments[0])&&value.arguments[0].elements.every(v=>ts.isStringLiteral(v));
ensure(valid,'compiled_guard_shape_unsupported');constants.set(name,`const ${name}=${value.getText(file)};`);
}
}
const context=createContext({setInterval(){throw Error('unexpected_scheduled_work');},clearInterval(){}},{codeGeneration:{strings:false,wasm:false}});
const body=functions[0].getText(file).replace(/^export\s+/,'');
return new Script([...constants.values(),body,'createMemoryPressureGuard;'].join('\n')).runInContext(context,{timeout:1000});
return new Script([...constants.values(),...helpers,body,'createMemoryPressureGuard;'].join('\n')).runInContext(context,{timeout:1000});
}
export async function verifyRecoveryBundle({bundle,manifest,expectedVersion}={}){
const profile=manifest?.agentmemoryCompatibility;
Expand All @@ -38,7 +45,7 @@ export async function verifyRecoveryBundle({bundle,manifest,expectedVersion}={})
let now=1000,available=8*1024**3,reads=0,resumes=0,pauses=0,releasePause;
const pausePending=new Promise(r=>{releasePause=r;});
const control={snapshot:()=>({status:'paused',pausedRequested:true,holds:['memory'],recovery:{attempts:2,maxAttempts:3,nextAttemptAt:0}}),change(action){if(action==='resume'){resumes++;return Promise.resolve(this.snapshot());}pauses++;return pausePending;}};
const guard=createGuard({control,schedule:false,clock:()=>now,readSample:async()=>{reads++;return{total:16*1024**3,available};}});
const guard=createGuard({control,schedule:false,clock:()=>now,monotonicClock:()=>now,readSample:async()=>{reads++;return{total:16*1024**3,available};}});
try{
ensure(guard.snapshot().automaticMemoryRecovery===false,'compiled_recovery_not_disabled');
for(let i=0;i<80;i++){now+=5000;await guard.refresh();await guard.drain();}
Expand Down
Loading
Loading