MuscleMemory Gives AI Coding Agents Reusable Procedural Memory

Summary

MuscleMemory is an open-source procedural-memory tool for AI coding agents, designed to reduce repeated exploration across sessions. It observes agent work and learns only after a task succeeds, using the final passing test as evidence; users must approve a skill before it can be shown to an agent. A learned note can include files to change, the test that validates the result, and warnings from failed attempts. In a benchmark using real Codex sessions, the project reports 25.5% fewer input tokens, 30.4% fewer tool calls, and 28.2% less time while all tasks still passed. A small Claude Code run using skills learned by Codex reported 43.5% fewer input tokens and 30% fewer tool calls, while a small SWE-bench Django run found that Codex solved as many issues with about half the tokens when using a related learned skill. The tool supports Codex, Claude Code, Cursor, CleanSlate, MCP clients, and Python or TypeScript SDKs. It runs locally by default with SQLite and a small on-device model, requires no account or outside calls, and can use PostgreSQL with pgvector for teams. The project says it is early: benchmarks cover repeat tasks on three controlled codebases, the Claude Code run is small, and cross-repository recall is untested.