Cognition adds Memory and daily Dreaming to Devin, open-sources the workflow

Cognition adds Memory and daily Dreaming to Devin, open-sources the workflow

N
News Editor
2026-10-06 13:46:24
Cognition has rolled out two new features for Devin: Memory and Dreaming. Memory lets Devin retain user preferences, corrections, and project-specific experience across sessions, while Dreaming runs as a daily background process that revisits and reorganizes those stored records. Users can inspect what Devin has remembered and review recent changes made by Dreaming. The system stores those records in a personal Memory Drive, which Cognition describes as a persistent Git repository organized mainly through Markdown files. A short MEMORY.md file serves as the entry point and index, and it is loaded into context at the start of each new session. Devin then searches only the relevant memory files for a given task, rather than loading the entire memory base into the prompt. Cognition said multiple Devin sessions can read from and write to the same memory set at the same time, with each session using its own Git checkout before changes are committed and merged back. The company has also open-sourced the setup as Agent Memory Repo. Cognition noted that while other agents such as Claude Code and Cursor can use a similar file-and-Git approach, Devin’s automatically scheduled daily Dreaming still requires extra configuration in the open-source version.

Cognition has introduced Memory and Dreaming for Devin, giving the coding agent a way to retain and reorganize information over time.

Memory is designed to help Devin carry user preferences, corrections, and project experience across sessions. Dreaming, by contrast, runs daily in the background to revisit and tidy up those records. Users can directly inspect what Devin has remembered and can also see what Dreaming changed most recently.

Memory Drive uses a persistent Git repository

The records are stored in a personal Memory Drive, which is essentially a persistent Git repository. Most of the content is organized as Markdown files. A short MEMORY.md file acts as the entry point and index, and it is placed into context at the start of each new session.

Devin then searches the relevant memory files based on the task at hand, in a way similar to searching a codebase. That means the full memory store does not need to be loaded into the prompt.

Multiple sessions can work on the same memory set

Multiple Devin sessions can read and write the same pool of memory at the same time. Each session has its own Git checkout. After a session updates memory, the changes are committed and merged back into Memory Drive. If concurrent edits happen, Git exposes the conflict so one session does not overwrite another.

Dreaming revisits, removes, and adds records every day

Dreaming itself runs as a background session each day. It reviews past conversations and existing memory, merges duplicate entries, removes temporary information or material that has not been used again for a long time, and adds experience that was not recorded earlier.

Cognition released the setup as Agent Memory Repo

Cognition has also separated the system into an open-source project called Agent Memory Repo. Other agents, including Claude Code and Cursor, can use a similar approach by storing long-term memory as files and managing them through Git.

The open-source release provides the memory structure and workflow. A daily auto-running Dreaming process like the one used by Devin still needs additional configuration.

This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan.
200

Disclaimer:

The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.

Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.