Cognition published a spec that lets AI coding agents keep their memory in a git repo of Markdown files
Cognition, the company behind Devin, released Agent Memory Repo on 4 October 2026, an MIT-licensed open spec that stores an AI coding agent's long-term memory as a git repository of Markdown files with cross-links and metadata, and the GitHub spec repo has 274 stars.
Image: Cognition
Why it mattersA team can see what their coding agent has learned, edit it by hand, review changes in a pull request and let two agents work from the same memory, which turns a hidden model-side state into a plain file every developer already knows how to read.
An AI coding agent that forgets what it was told yesterday is less useful by Friday than it was on Monday, and the memory it keeps today is usually a single file the user cannot read without the agent in front of them. Cognition, the company behind Devin, has published a spec for keeping that memory as a git repo of Markdown files instead.
Cognition released Agent Memory Repo on 4 October 2026 at cognition.com/agent-memory-repo, with the spec on GitHub at AgentMemoryRepo/agentmemoryrepo under the MIT license. The GitHub repository had 274 stars and 10 forks two days later. The company states on the page that the standard is open, and the GitHub README states that it was "originally developed by Cognition and released as an open standard."
The shape it defines
The entry point is a short MEMORY.md with cross-links written as [[path]], pointing at files organised in folders a team already knows how to arrange: projects/, billing/, one folder per domain. Each entry can carry metadata at the end of a line in the form [source: URL; added: YYYY-MM-DD], where source is a link back to the agent session that produced the fact and added is the date it was recorded. The repo can hold Markdown notes, SQL queries, scripts, or anything else git tracks.
Cognition describes an agent loop in four steps: clone the latest memory, search or follow a link to the relevant entries, update the entries as it learns, then commit after each edit. A second agent, called the Dreaming agent, runs periodically to look across sessions, pull out patterns and save them as new entries, and merge duplicates or resolve contradictions when it finds them.
What is new about it
The property Cognition leads with is composability. Two users can load their own memory repos into one agent session at the same time, each with its own ownership and git history, and the agent writes updates back to the right repo. Cognition's page gives the example of memory-alice/ and memory-bob/ sitting side by side on the same virtual machine. git handles the conflict resolution the standard way.
The second property is that the memory is a human-readable file. A team can review it, edit it by hand, open a pull request against it, and keep it in the same system their code lives in.
The spec repository also ships an installable skill, which the README lists Devin and Claude-compatible agents as the two targets for. No other agent SDK is named in the README, so getting the spec into Cursor, Codex or an in-house agent is a job the team would do themselves against the published format.
Source
- Primary: Cognition, Agent Memory Repo
- Spec and skill: AgentMemoryRepo/agentmemoryrepo on GitHub, MIT, 274 stars at the time of writing
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.