“The assistant should remember me” sounds like a feature request. The hard question is what it should remember, for how long, and how a person can correct it.
Memory is not a bigger context window. It is a product policy about which details remain useful after the current task ends.
Separate the scratchpad from long-term memory
An agent may need temporary notes while it completes a task: the selected customer, the current draft, or the remaining steps. Those details belong to the task session and can be discarded when the work is done.
Longer-lived memory should be smaller and more deliberate. A preferred report format may be useful next week; a one-time shipping address may not be. Store only information with a clear purpose, and make its source and last-updated time available to the system that uses it.
Keep facts, instructions, and permissions distinct
A note retrieved from a document is not automatically a trusted instruction. A customer email might contain useful facts, but it should not be allowed to rewrite the agent's permissions. Keep system rules, user preferences, and external content in distinct channels, and apply the same access controls to memory as to the original data.
This matters when the agent works across people or organizations. A remembered detail must not leak into another person's session simply because both conversations use the same assistant. Scope memory to the correct user and tenant, and check access when it is read—not only when it is stored.
Give people a way to inspect and forget
Users need a practical way to see what the product has retained, correct an outdated preference, and remove information. If an agent uses a remembered item to make an important suggestion, show enough context for the user to understand why.
Evaluate memory with realistic changes: a preference is updated, a once-relevant fact becomes stale, a user asks the system to forget something, or retrieved text contains misleading directions. Useful memory improves continuity without becoming invisible authority.
The best memory is not the most complete one. It is the smallest, safest set of context that makes the next interaction noticeably better.
