Learning Note: Agent Memory Is Not Just More Context
A short learning note about why agent memory should be selective, structured, and useful for future decisions.
What I Learned
Agent memory should not simply store everything the user says. If memory becomes a dumping ground, retrieval becomes noisy and the agent may use the wrong context.
A better memory system should answer three questions:
- What should be remembered?
- When should it be retrieved?
- How should it change future behavior?
My Current Understanding
I think useful memory has at least three layers:
- Profile memory: stable user preferences and long-term facts.
- Task memory: current goals, constraints, and progress.
- Reflection memory: lessons learned from previous failures or decisions.
Open Questions
- How can an agent decide whether something is worth remembering?
- How can we evaluate memory quality beyond retrieval accuracy?
- How can memory avoid reinforcing an early misunderstanding?