OKF Agent Memory: git-native, MCP context, 80% less token bloat
MIT, pure-Go, and built on Google's OKF standard: sub-300µs BM25 search, no database or embeddings, one MCP server for Claude Code, Cursor and Copilot.
Copy markdown
Your agent's memory lives in the repo now
OKF Agent Memory stores what a coding agent knows as plain Markdown-plus-YAML files in a knowledge/ directory, version-controlled with git — so you inspect, review and roll back its memory with git diff and git log instead of poking at an opaque vector store. It hit the Hacker News front page overnight.
80% fewer tokens, no DB, no embeddings
Local BM25 lexical search returns hits in under 300µs with a sub-15MB footprint and $0 per-query cost — no vector database and no embedding API to pay for. The project claims an 80% cut in token bloat versus re-stuffing context every turn. It's pure Go, MIT-licensed, with zero external dependencies.
Wire it into Claude Code or Cursor this weekend
`okf bootstrap` scaffolds a bundle in any project and `okf mcp knowledge` starts a native MCP server you point Claude Code, Claude Desktop or Cursor at. Agents query the local index before writing, so they stop duplicating notes and inventing divergent concepts across sessions.
It's Google's OKF standard, not a silo
The tool implements OKF v0.2, the vendor-neutral spec Google Cloud shipped July 24 that adds trust tiers (generated vs. verified), provenance via a sources field, and stale_after freshness dates. Your agent's memory stays portable across any OKF-aware tool rather than locked to one vendor's format.