Memory and dreaming: how Devin learns from working with you
By Steven Van ·
Memories are personal to each user in an organization, kept in a Git-backed drive, and a daily Dreaming session consolidates them and removes stale notes.
Devin now keeps Memory across sessions, along with a daily background process called Dreaming that tidies it. The announcement describes Memory as a way for Devin to carry over preferences, corrections and lessons about your projects and workflows. Memory is personal to each user within each organization on devin.ai, so it does not become shared instructions for the organization.
What Devin remembers
Memories are not session summaries. They are short notes on lessons Devin learned while working with you, each with a link back to the session where it was learned. Examples given include a correction to an assumption, a personal preference, an environment problem that took several attempts to fix, and a project decision that will matter again. Devin can retain these without you pausing the session to update a skill.
The memory drive
Notes live in a personal Memory Drive, a persistent Git repository of markdown files. They can be organized by repository, project or topic. A short MEMORY.md holds general preferences and an index of the other files.
- At the start of a session, Devin receives MEMORY.md as context. It then searches and reads relevant notes with the same tools it uses to navigate code, so the whole archive is not loaded into its prompt.
- Each session works on its own Git checkout of the drive. After editing a note, Devin commits, merges updates from other sessions and saves the result.
- If another session saves during a sync, a revision check rejects the stale write and Devin retries against the newer version. Conflicting edits are surfaced for resolution rather than silently overwritten.
Dreaming
Dreaming is a daily asynchronous session in which Devin reviews past conversations alongside its existing memory to improve the index. In a pass, Devin:
- consolidates overlapping notes
- removes transient details
- looks for useful lessons that were not captured during the original work
- removes stale memories and records that no session used
Source references and explicit preferences are preserved while the notes and index are reorganized.
Memory and skills
Skills capture a repeatable workflow, such as the steps for testing and publishing a change. Memory captures what Devin learns while working with you, such as tech stack preferences and the reason you gave. Skills are deliberately packaged for reuse, while memory accumulates through sessions and is revisited through Dreaming.
Where to find it
Memories and recent Dreaming sessions can be browsed on devin.ai under Customize → Memory. Cognition has also open sourced the standard it used to build the memory system, at cognition.ai/agent-memory-repo, and invites contributions and integration into other agents.
