okf-memory/okf-agent-memory

What it solves

AI agents often lose critical architectural decisions, domain knowledge, and operational facts once their context window closes. This project provides a persistent, vendor-neutral memory layer that allows agents to store and retrieve knowledge without relying on expensive, black-box vector databases or unstructured markdown files.

How it works

It implements the Open Knowledge Format (OKF) v0.2, storing knowledge as plain Markdown files with YAML frontmatter directly in a project's repository (within a knowledge/ folder). It uses a zero-dependency Go toolchain to provide fast, local lexical search (BM25) and graph validation. To integrate with AI agents, it includes a built-in Model Context Protocol (MCP) server, allowing agents to query and update their own persistent memory.

Who it’s for

Developers and AI agent users who want a version-controlled, transparent, and low-latency memory system for their agents across various domains, including software engineering, scientific research, and operations.

Highlights

• Git-Native: All memory is stored as plain text, making it auditable via git diff and git log with no external database required.
• High Performance: Features sub-300µs search latency and sub-5ms CLI startup times using a compiled Go binary.
• Zero API Costs: Uses local BM25 indexing instead of embedding APIs, eliminating recurring costs for memory retrieval.
• MCP Ready: Includes a native Model Context Protocol server for seamless connection to platforms like Claude Code and Cursor.
• Progressive Disclosure: Uses hierarchical index files and link graphs to prevent context bloat by loading only necessary concepts.