What should an AI assistant remember about you to be useful — and what should it never remember,…

An assistant should remember only what is stable, useful, and either explicitly provided for that purpose or easy for the user to approve.

Reasonable memory includes preferred language and format, accessibility needs, recurring project conventions, and the current state of a long-running project. I would not store passwords, API keys, payment details, private messages, medical or legal records, or sensitive political and religious inferences by default. It should also avoid silently inferring personality, income, health, or identity traits from casual conversation.

The design requirements matter as much as the list:

• show what is stored and where it came from;
• let the user inspect, edit, delete, export, and expire memories;
• attach purpose and confidence;
• require confirmation before using sensitive memory in a consequential action;
• do not use one project’s private context in another without permission.

The EU’s GDPR principles provide a useful baseline: purpose limitation and data minimisation mean collecting only what is adequate, relevant, and necessary for the stated purpose. Recent experiments on AI memory also found that relevant memory cues can help trust while intrusive references can damage it; editable controls improved trust on average but did not eliminate the penalty for inappropriate memory. See the study.

Convenience is not a sufficient reason to remember something forever. A good assistant should have a short memory by default and a deliberate, reversible memory when the user asks for more.