OKF Agent Memory ships v0.1.0 as git-native memory layer

What happened

OKF Agent Memory reached v0.1.0 with an initial commit on 5 September 2026. The project is a domain-neutral, git-native memory layer for AI coding agents, built on Google's Open Knowledge Format (OKF) v0.2. Within a day it collected 337 stars, 15 forks, and a Hacker News thread with 75 points and 22 comments.

OKF Agent Memory is a Go tool that stores an AI coding agent's persistent project memory as plain Markdown files with YAML frontmatter in a repository's knowledge/ folder, instead of a vector database. It ships a zero-dependency CLI, an embedded MCP server (okf mcp), and claims sub-300-microsecond BM25 concept search at zero embedding API cost.

What changed

• Memory lives in knowledge/index.md, log.md, and topic folders, versioned in git rather than a hosted database. A single Go binary (bin/okf) handles validate, search, show, create, update, init, and bootstrap, plus okf mcp to run the tool as a Model Context Protocol server over stdio.
• The project's own benchmark table claims concept search under 300 microseconds and full corpus parse plus graph validation around 4.0ms, against a stated 150ms-800ms for "Python / Vector DB Runtimes (Mem0, Letta)" and 40ms-120ms for Deno/Node.js tooling. It also claims zero dollars in retrieval cost per 1,000 queries, versus roughly $0.10-$0.50 for embedding-token approaches, and a memory footprint under 15MB.
• A follow-up commit on 6 September 2026 added MCP JSON-RPC 2.0 compliance and dynamic bundle resolution.
• The project is MIT licensed and describes itself as domain-neutral, with example bundles for software architecture, executive coaching, and literature review.

Why it matters for agent builders

Agent memory today mostly means unstructured files like CLAUDE.md or AGENTS.md, or a vector database that behaves like a black box. OKF Agent Memory sits between the two: it is structured enough to validate and search, but stored as text a developer can git diff, review, and audit like any other change. Its "search-before-write" convention is meant to stop an agent from duplicating or contradicting a concept it already recorded. Because the CLI embeds an MCP server, it connects to Claude Code, Cursor, and Codex without a separate integration layer. It also gives OKF v0.2's trust fields, provenance, generated-versus-verified status, lifecycle state, and freshness, described in Google Cloud's write-up of the standard, a working implementation at the tooling layer rather than just a specification.

What to watch

The Hacker News thread raised the open questions directly. Commenter langs asked why the project benchmarks latency instead of recall or precision, since millisecond gains matter little next to an LLM call, and esafak said they would like to see any tool in this space tested for accuracy rather than speed. Separately, mbreese and triyambakam questioned whether an agent will reliably choose a third-party memory tool over a harness's own built-in memory, pointing to Codex and Claude Code as the test cases. Commenters also flagged nearby projects worth comparing against, including huggingface/funes, mempalace/mempalace, fellowgeek/mcp-memory, and scaccogatto/okf-skills.

Key takeaways

• OKF Agent Memory v0.1.0 stores agent memory as Markdown with YAML frontmatter in a repository's knowledge/ folder, built on Google's OKF v0.2 standard.
• It ships a zero-dependency Go CLI and an embedded MCP server, connecting to Claude Code, Cursor, and Codex.
• The project claims sub-300-microsecond BM25 concept search, about 4ms graph validation, zero embedding API cost, and under 15MB memory footprint, against a comparison table that cites 150ms-800ms and paid embedding calls for tools like Mem0 and Letta.
• It reached 337 stars, 15 forks, and 75 points with 22 comments on Hacker News within a day of its 5 September 2026 initial commit.
• Hacker News commenters questioned the choice to benchmark latency over recall/precision and whether agents will actually adopt a third-party memory tool over a harness's built-in one.

FAQ

What is OKF Agent Memory built on?

OKF Agent Memory implements Google's Open Knowledge Format (OKF) v0.2, the standard Google Cloud describes as adding provenance, trust tiers, and lifecycle metadata to plain Markdown-and-YAML knowledge bundles.

How does OKF Agent Memory connect to coding agents?

The project's CLI includes an embedded Model Context Protocol server, started with okf mcp, which runs over stdio so agent platforms including Claude Code, Cursor, and Codex can call it directly to search, read, and write memory concepts.

Is OKF Agent Memory free to use?

The project is MIT licensed, and because concept search runs on a local in-memory BM25 index rather than an embedding API, the README states retrieval carries no per-query API cost.

Source: okf-memory/okf-agent-memory Discussion: Hacker News

Tools like Mem0 that OKF Agent Memory benchmarks itself against are tracked alongside other memory and coding tools in the MCP servers directory on AI Agents Listing.