OKF Agent Memory – Git-native persistent memory for AI coding agents

OKF Agent Memory addresses the problem that AI agent conversations reset when context windows close, losing architectural decisions and domain facts. Instead of unstructured files like CLAUDE.md or black-box vector databases, it stores project memory in plain Markdown/YAML under a knowledge/ directory directly in the repository. The design builds on the Open Knowledge Format (OKF) v0.2 specification, adding a behavioral "Agent Memory Convention" with rules for search/review/trust, plus LLM prompt skills and a deterministic Go library/CLI that supports parsing, validation, search, and a Model Context Protocol (MCP) server.

Performance is emphasized: in-memory BM25 search runs in under 300 microseconds, full corpus parse/graph validation in about 4 ms for 50+ concepts, and CLI cold start under 4 ms — all from a single zero-dependency Go binary. This avoids vector embedding API costs and network roundtrips, and everything is version-controlled plain text that can be audited with git diff/log. The toolkit supports provenance (sources), trust tiers (generated vs. verified), lifecycle metadata (status, stale_after), and progressive disclosure via hierarchical index files so agents load only needed concepts. The "search-before-write" principle reduces duplication and hallucination divergence. Benchmarks compare OKF's microsecond latency to 150–800 ms for Python/vector DB runtimes (Mem0, Letta) and 40–120 ms for Deno/Node.js tooling.

This is a truly domain-neutral approach targeting software engineering, coaching, scientific research, literature review, and operations. By keeping memory in-repo and open-standard, it aims to become a standardized memory layer for AI agents without lock-in, with practical limits tested in agent workflows.