What should an agent remember?

My starting question for agent memory is simple: what future task does this information help with?

A standing communication preference might matter across many tasks. A temporary campaign instruction may matter for a week. A discarded idea may have no future purpose at all. Treating them alike makes it harder to know which context should guide the next answer.

I would separate the current task from longer-lived context. For each durable item, I would want to know where it came from, when it was last confirmed, and what scope it applies to. A project decision should not silently become a rule for every other project.

Correction matters as much as storage. People change their minds. If someone gives a new instruction, the system should be able to recognize that an older preference may no longer apply. The person should also have a practical way to inspect, correct, or remove remembered information.

For a prototype, I would test a few simple sequences. Save a preference, override it, and ask a related question. Save a temporary fact, let it expire, and see whether it still influences the response. Add information to one project and check whether it leaks into another project's answer.

Those tests would tell me more about the behavior than the size of the memory store.

The memory system I want would make context useful and accountable. It would remember enough to help, while letting a person understand why that information is still there.