Git-native MCP server for auditable agent memory and fast retrieval
okf-agent-memory by Okf Memory is an MCP server that provides persistent, auditable memory for AI coding agents. The tool maps project knowledge into a Git-based corpus and offers programmatic retrieval and writing endpoints so agents can read and record task context without opaque backends. It stores memory as human-readable files inside the repository and exposes a protocol surface for MCP clients. Intended for AI developers and engineering teams building agentic workflows who need traceable, repository-integrated project memory.
What tasks can you actually use it for?
The tool implements the Open Knowledge Format v0.2 to convert a repository into a self-validating knowledge corpus, so agentic processes can persist structured notes alongside source files. This design replaces RAG approaches that rely on separate vector stores by keeping memory as readable project artifacts, making an agent's reasoning traces auditable and manageable with the project's existing review and history tools.
How accurate are the retrievals compared to doing it manually?
The tool runs an in-memory BM25 lexical index measured at sub-300 microsecond ranking across titles, metadata, tags, and bodies, enabling near-instant local lookup without network latency. Progressive disclosure returns compact, task-focused slices, typically around 300 tokens, and the project states selective retrieval can reduce prompt context bloat by up to 90 percent for coding tasks.
What file formats and storage layout does it require?
The system expects human-readable Markdown files placed under a knowledge directory inside a Git repository and uses YAML metadata to mark trust tiers that separate human-verified entries from agent drafts. Because memory is plain text under source control, standard Git commands such as diff and blame work for reviewing and reverting memory changes, fitting the same review workflow used for code.
Does it require external services or special infrastructure?
The app ships as a single, zero-dependency Go binary with sub-4 millisecond cold starts and runs on macOS, Linux, and Windows. It exposes MCP actions such as okf_search, okf_show, and okf_create to compatible clients including Claude Desktop and Cursor. Search and retrieval operate locally without embedding API calls, which the developer describes as deterministic and offline, avoiding runtime LLM calls for lookups.
Best for engineering teams that accept repository-first agent memory
The app suits AI developers and engineering teams who prioritize auditable, local project context and who can incorporate memory into their source control workflows. Expect trade-offs if your use case depends on embedding-based semantic similarity rather than lexical ranking. Pair the tool with normal code review practices to validate agent-generated drafts and keep project memory aligned with human-reviewed knowledge.
