Right now the main agent doesn't have much guidance on when to reach for the memory tool. We need to update the system prompt so it knows to call the memory subagent in situations like:
- User states a preference ("I prefer tabs over spaces")
- User repeatedly corrects the same behavior
- Agent discovers a project convention during a task
- A workflow pattern emerges that would be useful next session
We also need to make sure it doesn't over-save - transient task context, scratch notes, or things that only matter for the current session shouldn't end up in memory.
Needs at least one eval showing the agent proactively saves something it should, and doesn't save something it shouldn't.
Right now the main agent doesn't have much guidance on when to reach for the memory tool. We need to update the system prompt so it knows to call the memory subagent in situations like:
We also need to make sure it doesn't over-save - transient task context, scratch notes, or things that only matter for the current session shouldn't end up in memory.
Needs at least one eval showing the agent proactively saves something it should, and doesn't save something it shouldn't.