A memory format for AI agents skips vector databases entirely

OKF Agent Memory stores project memory as Markdown files with YAML frontmatter, then searches them with in-memory BM25. It skips vector databases entirely. It bets git-native storage beats a database for most coding agents. The project says that cuts token use by 80%, since the agent gets summaries before full file contents.

Most agent setups pick one of two extremes. A single CLAUDE.md file grows until nobody can read it. A vector store turns memory into a black box nobody can diff. Markdown with YAML frontmatter sits between them. You can open a memory file in a text editor. You can put the whole knowledge/ directory under version control the same way you'd track source code.

The project ships a Go benchmark runner. Anyone can reproduce the token-reduction number locally against LM Studio or Ollama instead of taking it on faith.

OKF Agent Memory implements Open Knowledge Format v0.2 underneath, a spec Google published inside GoogleCloudPlatform's knowledge-catalog repo, which has 9,079 stars. That kind of backing could give agent frameworks a shared memory schema rather than each one inventing its own.

The repo doesn't compare BM25 search against a vector-backed setup on the same corpus size. Token savings are believable for one project's docs. Whether flat-file search still finds the right document once a knowledge base spans hundreds of files is a question the current numbers don't answer.