Persistent memory for AI coding agents.
Powered by iii-engine.
Quick Start • Why • Agents • How It Works • Search • Memory Evolution • MCP • Viewer • Configuration • API
Every AI coding agent has the same blind spot. Session ends, memory vanishes. You re-explain architecture. You re-discover bugs. You re-teach preferences. Built-in memory files like CLAUDE.md and .cursorrules are 200-line sticky notes that overflow and go stale. agentmemory replaces that with a searchable, versioned, cross-agent database — 41 MCP tools, triple-stream retrieval (BM25 + vector + knowledge graph), 4-tier memory consolidation, provenance-tracked citations, and cascading staleness so retired facts never pollute your context again. One instance serves Claude Code, Cursor, Codex, Windsurf, and any MCP client simultaneously. 581 tests. Zero external DB dependencies.
The result is measurable. On 240 real observations across 30 sessions, agentmemory hits 64% Recall@10 and perfect MRR while using 92% fewer tokens than dumping everything into context. When an agent searches "database performance optimization," it finds the N+1 fix you made three weeks ago — something keyword grep literally cannot do. Memories version automatically, supersede each other, propagate staleness to related graph nodes, and sync across agent instances via P2P mesh. Your agents stop repeating mistakes. Your context stays clean. Your sessions start fast.
npx @agentmemory/agentmemory # installs iii-engine if missing, starts everything
Why agentmemory
AI coding agents forget everything between sessions. You explain the same architecture, re-discover the same patterns, and re-learn the same preferences every time. agentmemory fixes that.
Session 1: "Add auth to the API"
Agent writes code, runs tests, fixes bugs
agentmemory silently captures every tool use
Session ends -> observations compressed into structured memory
Session 2: "Now add rate limiting"
agentmemory injects context from Session 1:
- Auth uses JWT middleware in src/middleware/auth.ts
- Tests in test/auth.test.ts cover token validation
- Decision: chose jose over jsonwebtoken for Edge compatibility
Agent starts with full project awareness
No manual notes. No copy-pasting. The agent just knows.
What it gives you
| Capability | What it does |
|---|---|
| Automatic capture | Every tool use, file edit, test run, and error is silently recorded via hooks |
| LLM compression | Raw observations are compressed into structured facts, concepts, and narratives |
| Context injection | Past knowledge is injected at session start within a configurable token budget |
| Semantic search | Hybrid BM25 + vector search finds relevant memories even with different wording |
| Memory evolution | Memories version over time, supersede each other, and form relationship graphs |
| Project profiles | Aggregated per-project intelligence: top concepts, files, conventions, common errors |
| Auto-forgetting | TTL expiry, contradiction detection, and importance-based eviction keep memory clean |
| Privacy first | API keys, secrets, and <private> tags are stripped before anything is stored |
| Self-healing | Circuit breaker, provider fallback chain, self-correcting LLM output, health monitoring |
| Claude Code bridge | Bi-directional sync with ~/.claude/projects/*/memory/MEMORY.md |
| Cross-agent MCP | Standalone MCP server for Cursor, Codex, Gemini CLI, Windsurf, any MCP client |
| Citation provenance | JIT verification traces any memory back to source observations and sessions |
| Cascading staleness | Superseded memories auto-flag related graph nodes, edges, and siblings as stale |
| Knowledge graph | Entity extraction + BFS traversal across files, functions, concepts, errors |
| 4-tier memory | Working → episodic → semantic → procedural consolidation with strength decay |
| Team memory | Namespaced shared + private memory across team members |
| Governance | Edit, delete, bulk-delete, and audit trail for all memory operations |
| Git snapshots | Version, rollback, and diff memory state via git commits |