OKF Agent Memory
Local AI agent memory stores project knowledge in Git, connects to Claude Code, Cursor and Codex through MCP, and uses an MIT license.
OKF Agent Memory gives AI coding agents persistent project knowledge stored locally as plain text in Git. It's for developers who want architectural decisions and project constraints to survive across sessions without putting the entire knowledge base into every prompt. The project is open source under the MIT license.
Its memory model separates the rules an agent receives at session start from the facts it retrieves when needed. Compact instructions keep project guardrails in context, while linked knowledge files let the agent load relevant details on demand. Memory search runs locally using BM25, without an external database or embedding API.
An embedded MCP server connects the memory store to Claude Code, Cursor and Codex. The Go tool also provides a command-line interface and library. Because the knowledge lives in version-controlled files, developers can inspect edits, review proposed changes and trace their history alongside code.
The Open Knowledge Format records sources, distinguishes generated knowledge from verified facts, and tracks status and staleness. Validation checks the knowledge graph and description drift. Decisions can link directly to source files, so an agent can retrieve the constraints that apply before editing them. A search-before-write convention helps avoid duplicate concepts.
Local storage and retrieval don't require network calls. Optional Hub sync sends encrypted knowledge to a remote service using a zero-knowledge vault.
