AgentMemory · 2026-05-13 · notable
AgentMemory v0.9.12 — Persistent Memory Layer for Claude Code, Cursor, and Gemini CLI Hits GitHub Trending #3 With 8.7K Stars, Hybrid BM25 + Vector + Knowledge-Graph Search Over MCP
Apache-2.0 TypeScript service captures every PreToolUse/PostToolUse/SessionStart hook, indexes them across BM25 + vectors + a temporal knowledge graph, and serves recall to coding agents over MCP and REST.
A drop-in persistent memory service for coding agents — Claude Code, Cursor, Gemini CLI — with BM25, vector, and knowledge-graph retrieval over MCP.
Key specs
| License | Apache-2.0 |
|---|---|
| GitHub stars | 8,666 |
| Language | TypeScript |
| Mcp tools | 51 |
| Rest endpoints | 121 |
| Hooks | 12 |
| Long mem eval retrieval r5 | 95.2% |
| Token savings vs full context | 92% |
| P50 latency ms | 20 |
What is it?
AgentMemory is an Apache-2.0 TypeScript project that gives AI coding agents a long-term memory store across sessions and across tools. Instead of pasting CLAUDE.md or notepad blobs every time, agents query a local service that has been quietly indexing tool calls, file edits, and session metadata. v0.9.12 shipped on May 13 and the project is at #3 on GitHub trending today.
How does it work?
Twelve hooks (PreToolUse, PostToolUse, SessionStart, Stop, and friends) feed raw observations into a local SQLite store. A background pipeline consolidates them across four tiers — working, episodic, semantic, procedural — and indexes the result three ways: BM25 for keyword recall, vector embeddings for semantic recall, and a temporal knowledge graph for entity-edge queries. Agents pull top-K context through 51 MCP tools or 121 REST endpoints; a viewer dashboard runs locally on port 3113.
Why does it matter?
Agent memory is fragmenting into a half-dozen approaches (Anthropic Managed Agents Memory, Cloudflare Agent Memory, Letta Code, Hippo Memory, OpenChronicle). AgentMemory bets on running everything on-device with no external database, while staying compatible with most coding-agent clients via MCP. v0.9.12 hardens unicode tokenization for non-Latin scripts, plugs an HTTP-token leak, and stops the viewer crashing on init errors.
Who is it for?
Developers running long-lived sessions across Claude Code, Cursor, Codex CLI, Gemini CLI, or Windsurf who want persistent context without paying for a hosted memory service.
Try it
npx @agentmemory/agentmemory