OKF Agent Memory provides a domain-neutral, Git-native persistent memory layer for AI coding agents based on Google OKF v0.2. It stores knowledge directly in repositories as plain Markdown files with YAML frontmatter, bridging ad-hoc markdown files and black-box vector databases.

The system implements five layers: OKF v0.2 specification, agent memory convention, agent skills, Go-based tooling, and the project knowledge corpus. Key highlights include sub-300µs in-memory BM25 search, ~4ms graph validation, zero vendor lock-in, and zero API costs for retrieval.

Performance benchmarks show significant advantages over Python/vector DB runtimes: 150ms-800ms search latency drops to under 300µs, cold-start overhead falls below 4ms, and memory footprint stays under 15MB. Retrieval costs are zero compared to $0.10-$0.50 per 1,000 queries.

Built in pure Go with zero external dependencies, the tool offers a Model Context Protocol server, progressive disclosure to reduce context bloat, and search-before-write principles. It supports software engineering, coaching, scientific research, and operations domains. Users can reproduce benchmarks locally with LM Studio or Ollama.