The Memory Contract: Personalization Without Losing Control

Memory can turn a generic assistant into a useful collaborator. It can also preserve an old mistake, expose private context in the wrong place, or make the system feel as if it knows more than the user intended.

The difference is a clear memory contract.

Memory Needs Scope

Not every remembered fact belongs everywhere. Separate personal preferences, project knowledge, organizational policy, and temporary task state. A writing preference may travel across documents; a confidential project detail should not.

Each memory should carry:

• Its source and the time it was recorded.
• The scope where it may be used.
• A confidence or verification status.
• A retention rule.
• A path for correction and deletion.

Inferred preferences deserve extra caution. “The user chose this once” is not the same as “the user always wants this.”

Make Memory Legible

Users should be able to inspect the memories affecting a response, disable a category, correct a detail, and start a task with a clean context. The system should explain personalization in normal language rather than revealing an internal database.

For sensitive workflows, use an explicit save action instead of automatic memory. For shared devices or delegated work, make the active identity and memory scope visible before the conversation begins.

Forgetting Is a Feature

Some information should expire naturally. Temporary travel plans, a one-off tone request, or an outdated team structure can become harmful if retained indefinitely.

Useful personalization is not maximal recall. It is selective continuity under the user’s control.