Vibeleaderboard
← All Intel
Intel / article

Cognition's Agent Memory Repo: git-backed memory with a Dreaming agent

Source
cognition.com
Author
Cognition
Date

Context

Cognition's Memory Repo treats an AI agent's memory as an ordinary Git repository of files, so the tools developers already use for code also handle history, sharing and conflicting edits. The specification lays out a fixed loop for every session: clone the latest memory, search it or follow links for what the task needs, update entries as the agent learns without a person approving each change, and push after every edit. A short MEMORY.md file is the entry point that the agent loads at the start of each session, with links to more detailed notes, saved SQL queries or scripts. Inside a note, each entry is a one-line bullet that can carry a source link back to the session where the fact was learned and the date it was added.

Dreaming is meant to keep that memory from going stale. In its announcement, Cognition says its Devin agent builds a memory of how a user likes to work and improves it overnight. The specification describes Dreaming as a separate agent that runs periodically with two jobs: spot patterns across sessions and save them as new entries, and clean up by merging duplicates, removing outdated entries and checking sources to settle contradictions. The cleanup job is aimed at an obvious weakness of a memory that every session writes to: left alone, it keeps growing and keeps stale facts beside newer ones.

Because memories are folders, the format can combine them. When a second person joins a session and chooses to share, their repository is cloned beside the first, and the agent writes each person's preferences to their own repository, asking when the destination is unclear. Several agents working in parallel can push to one repository; when two of them edit the same line, Git rejects the second push and that agent reads both versions before writing one. The page's worked examples, including a four-agent swarm that traces slow checkout requests to a cache refresh, are illustrations with placeholder sources, not reported results. The installable trial is also narrower than the announcement: it keeps a local repository with no remote, runs only when the user invokes the , and does not include automatic startup or scheduled Dreaming. Cognition says it released the format as an open standard and accepts contributions on GitHub. The announcement's video could not be reviewed.

Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
Read the source cognition.com
More from Cognition
Recommended reads
Comments

Checking sign-in…

Loading comments…