AI agent memory that lives in git and searches in microseconds

What it does

OKF Agent Memory is a Go binary that turns a knowledge/ directory inside your repo into a structured, persistent memory bank for AI coding agents. It stores concepts as Markdown files with YAML frontmatter following Google’s OKF v0.2 spec, then serves them back to agents through an embedded Model Context Protocol server or a fast CLI. Everything stays in git—no external database, no embedding API calls, and no network roundtrips.

The interesting bit

Instead of dumping everything into the context window and hoping for the best, it uses progressive disclosure through hierarchical index.md files and link graphs so agents load only the concepts they actually need. Search is pure in-memory BM25, which keeps retrieval under 300 µs and the whole binary cold-start under 4 ms—fast enough to live inside tight agent tool-calling loops without the usual VM boot tax.

Key highlights

• Zero external dependencies; single Go binary with a memory footprint under 15 MB RSS.
• Sub-300 µs concept search and ~4 ms full corpus graph validation, per the project’s own benchmark table.
• Enforces a “search-before-write” convention to prevent duplicate concepts and hallucinated drift.
• Built-in MCP server over stdio plugs into Claude Code, Cursor, and Codex without extra tooling.
• Claims an 80% token reduction via progressive disclosure, with a reproducible benchmark suite included in the repo.

Verdict

Worth a look if you run local or self-hosted agents and want version-controlled, offline memory without spinning up a database or paying embedding fees. Skip it if you need the semantic similarity retrieval of a full vector database or prefer managed memory outside your repository.