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

The Open Knowledge Format (OKF) has introduced Agent Memory, a domain-neutral, Git-native persistent memory solution tailored for AI coding agents. The OKF Agent Memory addresses the issue of conversations with AI agents resetting once context windows close, resulting in the loss of valuable architectural decisions, domain discoveries, and operational facts.

To combat this, OKF Agent Memory provides a standardized, vendor-neutral memory layer that resides directly within your repository (knowledge/) as plain Markdown files with YAML frontmatter.

Built in Go with no external dependencies, the OKF Agent Memory is engineered for high-frequency agent tool calling loops. It offers an automated benchmark runner written in pure Go to verify Time-To-First-Token (TTFT) speedups and up to -80% token reduction on your local hardware. This can be achieved using LLMs such as LM Studio, Ollama with Gemma, Qwen, or Llama. To get started, simply clone the repository and compile the standalone okf executable located in the bin/ directory.

The OKF Agent Memory architecture can be easily integrated into any new or existing repository with a single command. It ships with a native Model Context Protocol (MCP) server over stdio, which seamlessly connects with various agent platforms like Claude Code, Cursor, Codex, and others. The repository includes a self-documenting knowledge bundle that implements Google OKF v0.2, featuring sub-300µs in-memory BM25 search, an embedded MCP server, and progressive disclosure.

By utilizing OKF Agent Memory, users can slash token bloat by 80% while avoiding the need for external databases or dependencies.