// the find
rohitg00/agentmemory
#1 Persistent memory for AI coding agents based on real-world benchmarks
agentmemory gives AI coding agents persistent memory across sessions via MCP, hooks, or REST API. It stores session observations locally (SQLite + the iii-engine runtime), then injects relevant context at the start of each new session. Aimed at developers who are tired of re-explaining their architecture, preferences, and past decisions every time they start a new Claude Code or Cursor session.
Zero external dependencies — everything runs on SQLite and the bundled iii-engine, so there's no Postgres or Qdrant to stand up. The hybrid retrieval (BM25 + vector + graph with RRF fusion) is meaningfully better than keyword search alone, and the LongMemEval-S numbers are reproducible from the eval harness in the repo. The 12 auto-capture hooks for Claude Code mean you don't have to manually call any save API — memory accumulates silently as you work. The 4-tier consolidation and decay pipeline means old memories don't just pile up forever; the system actually prunes and compresses over time.
The iii-engine dependency is a black box — agentmemory pins a specific version (v0.11.2) and explicitly refuses to attach to any other, which means you're locked to whatever that engine does and can't control your upgrade path independently. Windows support is second-class: the fast install path requires WSL2, the `agentmemory connect` command is unsupported on native Windows, and the README buries this after a long install section. The benchmark comparisons are partly apples-to-oranges — their own R@5 is on LongMemEval-S while competitors' numbers come from LoCoMo; the table presents them side-by-side without enough warning. The 25K-star count relative to what is essentially a local SQLite wrapper around an opaque engine runtime feels inflated by the trending cycle rather than reflecting adoption depth.