Skip to content

Memory ​

Project memory is what the team wants the agent to know about a repository every time: conventions, commands that work, pitfalls. It is versioned per project and per role, and only a version someone has explicitly published is used.

How a turn uses it ​

When a turn starts, it takes the latest published memory for its project and role (ask or code) and keeps it for the whole turn.

  • The agent gets the published memory with its instructions.
  • In Code turns, the memory is also a file in the sandbox, MEMORY.md. The agent may edit it to propose updates.
  • When the turn ends, any edits come back as memory-draft.md under "Raw logs and artifacts".

A draft never changes the team's memory by itself. Someone has to publish it.

The agent is told to treat memory as task data, like the repository's own files.

Publish a version ​

Owners and administrators open "Project memory" in the settings:

  1. "Select project memory" for the project.
  2. Click "Record project memory", choose the role and write or paste the content, for example from a turn's memory-draft.md. Up to 30,000 characters.
  3. Tick "Publish for use in future tasks" to make it the version new turns use, or leave it unticked to keep a draft.

Each save is a new version. The list shows every version as "Draft", "Published" or "Deleted". Turns already running keep the version they started with.

Delete a version ​

"Delete content" clears a version's text. Deleting the latest published version does not bring back an older one. New turns get no memory until someone publishes again.

Saving, publishing and deleting are recorded in the organization's audit log.