How agents learn
Agents learn from conversations and from the team's corrections. What an agent knows enters its context when relevant. On its Memory tab, edit, pin or archive what's wrong or stale.
Memory grants no permission
Nothing in memory lets an agent do more. What it can do stays in its Tools and access, and sensitive actions always ask for approval.
Proposed changes
When a lesson repeats, the agent may propose changing its own instructions or the workspace's context. Nothing changes until someone accepts, and the text can be adjusted first. The previous version stays in the history. A refused proposal is kept as a lesson, so it isn't proposed again.