deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

OKF Agent Memory keeps agent knowledge in git as plain Markdown files

OKF Agent Memory stores agent knowledge as Markdown files with YAML frontmatter in a git-tracked folder, adding provenance flags and on-demand retrieval over MCP so teams can review what agents remember.

OKF Agent Memory keeps agent knowledge in git as plain Markdown files

A new entry in agent persistence

AI agents usually begin every session from scratch. Unless someone deliberately recorded a decision, the reasoning behind last week's work disappears and the agent has to rediscover it. OKF Agent Memory, an open-source tool announced in a launch post on dev.to, attacks this problem by storing agent memory as ordinary Markdown files that a team can review the way it reviews code.

According to the announcement, version 0.8.0 shipped on 5 November 2026, the project is MIT-licensed, and its GitHub repository had 866 stars at the time of writing. The post also describes the tool as an implementation of Google's Open Knowledge Format v0.2; that specification claim comes from this single source, so teams evaluating the tool may want to verify it independently.

Markdown with YAML frontmatter, tracked in git

The central design decision is that memory does not live inside a proprietary database or hosted service. Entries are Markdown documents with YAML frontmatter, held in a knowledge folder under version control. Because the format is plain text and git-native, an agent's accumulated knowledge can go through pull requests, diffs and code review — processes engineering teams already have in place.

Every entry also carries provenance information stating whether it was written by an agent or verified by a human. The announcement recommends that a person confirm an entry before the agent treats its contents as established fact, which turns memory maintenance into a supervised workflow rather than silent accumulation.

A small working set, retrieved on demand

The tool does not push everything a team knows into the model's context. Instead, it maintains a working codex — the tool's term for a compact summary — capped at 400 tokens, while the complete knowledge bundle is fetched only when needed. Retrieval relies on BM25 search that the author reports finishing in under 300 microseconds, and the memory is served to agents over MCP (Model Context Protocol), so any client that speaks the protocol can consume it.

The performance figure is the author's own claim from a promotional launch post — the announcement also links to the author's newsletter — so the specifics deserve scepticism until independently confirmed.

What the author recommends

The post advises seeding the knowledge folder with recorded decisions and their rationale rather than transcripts of agent conversations. Agent-written entries then await human verification, keeping people in the loop on what eventually counts as organisational knowledge.

Why it matters

Persistence is one of the most practical gaps in current agent tooling. Agents that forget everything between sessions make inconsistent decisions, repeat earlier work and never build up team knowledge. Many existing approaches keep memory inside a vendor's service, where it is hard to audit, export or review with standard engineering tools.

Plain files in git invert that arrangement. Memory becomes subject to the same review, branching and history as code, provenance is explicit on every entry, and the combination of an MIT licence with MCP serving makes the tool straightforward to trial across different agent frameworks. It is still an early project at v0.8.0 with a single announcement as its source, but the underlying idea — memory as reviewable text rather than opaque internal state — is a sensible direction for teams whose workflows already run on git.

  • #agent-memory
  • #markdown
  • #mcp
  • #open-source
  • #developer-tools

Related posts