Memory sounds like a technical feature, but users experience it as product behavior. They notice when an agent remembers the right preference and when it confidently carries forward an outdated assumption.
The useful boundary
Good memory is selective. It stores durable preferences, recurring context, and decisions that reduce future work. It does not preserve every conversation forever.
Make memory legible
Users should be able to inspect, correct, and delete what an agent knows. The interface for memory is part of the trust model, not an administrative afterthought.
