mnemex
Local memory for AI coding agents that pins every decision to the code symbols it governs.
- Domain
- Agent memory
- Built with
- Python, MCP, SQLite FTS5
- Source
- github.com/notsointresting/mnemex
The problem
Agents remember decisions as sentences ("auth is stateless") but lose track of which code the decision covered and when that code changed.
How it works
- Indexes Python and TypeScript symbols, imports and calls into one local SQLite file.
- Anchors each decision to symbol content hashes and reports it as fresh, stale or orphaned.
- FTS5 and BM25 search for sub-millisecond retrieval, with 56× token compression over raw session history.
- MCP tools over stdio and local HTTP, plus a diagnostic CLI, mnemex doctor.
Where it landed
No network calls, no embedding model, no telemetry, and secrets redacted before storage. Runs on Python 3.10 to 3.13.
Stack
- Python
- MCP
- SQLite FTS5
- BM25
- AST parsing