Before you let it remember

1. 01The user can see everything the system remembers about them, in plain language, in one place.Legible memory means readable sentences, not a JSON export.
2. 02Every remembered item can be individually edited or deleted — and deletion actually takes effect.Test it. Delete a memory, then probe for it. Ghosts are betrayals.
3. 03The system tells the user at the moment it saves something new to memory.Silent accumulation is surveillance with a friendly font.
4. 04Personalized behavior is traceable: the user can find out why the system adapted.
5. 05Thumbs-down (or any feedback) produces a change the user can later observe.Feedback that visibly matters gets given; feedback into a void stops within a week.
6. 06Model or behavior updates that users will notice are announced before they land.Never surprise with change. A workflow that breaks silently reads as regression, not progress.
7. 07A returning user’s accumulated context transfers — they never re-explain what the product already knows.
8. 08There is a clean way to start fresh: reset memory without deleting the account.
9. 09We have decided what the system must never remember, and the list is enforced in code, not policy.Health, finances, credentials, off-hand venting — name the exclusions explicitly.
10. 10Someone owns the quality of accumulated memory six months out, by name.Stale memory is worse than no memory. Rot needs an owner.

Discussion prompts

• Would our users be comfortable reading their own memory file out loud in a meeting? If not, why are we comfortable storing it?
• What happens to a user’s trust the first time the product remembers something they didn’t know it heard?
• When we upgrade the model and outputs change tone overnight, what tells the user it was us and not them?
• Whose interests does this personalization serve when the user’s preference and our engagement metric disagree?
• What does this feature look like after two years of accumulated memory — sharper, or haunted by outdated context?
• If a user asked “why did it do that?” about an adapted behavior, how many of us could answer from the UI alone?