AI agents solve problems, forget what worked, and repeat the same work. MuscleMemory fixes that. It's an open-source procedural memory layer that turns successful agent runs into reusable, test-verified skills. Skills require repeated passing tests and human approval before reuse. Failed attempts become warnings, not trusted solutions. Works with Claude Code, Codex, Cursor, and MCP-compatible agents. In Benchmarks: 30% fewer tool calls, 25% fewer input tokens, with the same task success rate.