OKF Agent MemoryMCP Server

Most agent memory systems reach for a vector database and an embedding API. OKF Agent Memory takes the opposite position: memory is plain text in your git repository, searched lexically in process, and it costs nothing per query.

The design splits memory in two, which is the part that does the work. A compact behavioural codex — roughly 100–150 tokens — holds normative instructions and is always in context. Everything else lives in a searchable knowledge bundle that is retrieved on demand through progressive disclosure. That separation is what lets the project claim around 80% token reduction against a monolithic instructions file, because the agent no longer pays for domain knowledge it is not currently using.

Search is in-memory BM25 with sub-300µs latency and zero external databases, so there is nothing to run alongside the agent and nothing to pay an embedding provider for. Because every artefact is version-controlled plain text, memory diffs in review, travels with a branch, and carries no vendor lock-in. The MCP server is embedded in the same Go binary and works with Claude, Cursor and other compatible clients.

It implements Google's OKF v0.2, the open standard format for agent knowledge, including provenance metadata, trust tiers that distinguish generated from verified knowledge, and lifecycle tracking — so entries carry where they came from and how much they should be trusted. An Agent Action Grammar (AAG) gives instructions a deterministic, token-efficient syntax. Recent releases have concentrated on hardening the MCP server against input bounds violations, denial-of-service, Unicode spoofing and path traversal. MIT licensed; build with make build to produce a standalone bin/okf.