Active Memory
Review lessons that later runs can recall in your organization.
How lessons carry forward
Agents recall relevant lessons while working. After the graph settles, a postmortem can save or update lessons supported by the run's evidence; zero new lessons is valid. Lessons stay in your organization outside workspace checkpoint retention.
Repository | Advice matched to a repository and its files, folders, identifiers, or keywords. |
|---|---|
Graph | Advice tied to a graph and node, optionally qualified by a repository. |
Memory does not restore files, services, or conversations. Use checkpoints for workspace recovery.
Review or edit a lesson
Open Memory for a table of every lesson with its repository and kind. Search, or filter by Repository or Kind. Graph lessons without a repository show Any repository and appear only with no repository filter. Open a lesson for its recommendation, what it applies to, why a run saved it and the run it was learned from.
Members can read lessons. Owners and admins can edit a lesson's title and recommendation on its page, or delete it; what it applies to and why it was saved stay fixed. If the lesson changed while you were editing, your draft stays: compare it with the latest version, then use the latest version or keep your draft and save again.
Control recall and maintenance
Recall and postmortem default to enabled. The postmortem uses a source agent's model, runtime, and connections in separate sessions and can make additional model calls. Runtime settings can disable repository or graph memory independently, or disable postmortem while keeping recall. Follow the Active Memory configuration reference.
The MCP memory tool manages existing lessons under the same human permissions. API key delegations cannot access Memory.