Memory Needs a Front Door and an Exit
Memory should be a user-governed resource, not an invisible capability.
Personalised assistants promise to remember preferences, projects, relationships, and working styles. Memory can reduce repetition and make help more relevant. It can also make interaction opaque: why did the system assume this? Where did the detail come from? Who can see it? How can it be corrected? A memory feature should not be a hidden drawer that only the system can open.
Memory as a user-facing object
Imagine returning to a writing assistant after several weeks. It remembers that you prefer concise explanations and are working on a particular project. The interface should show which details are being used and why. A useful memory surface offers a front door to inspect what is remembered, source labels distinguishing explicit statements from inferred patterns, a correction path, scope controls, and an exit that deletes or disables a memory.
Inference is not a fact
There is a difference between “You asked me to use short examples” and “I noticed you often choose short examples.” The first is an explicit preference; the second is an inference. A system should not turn weak patterns into strong claims about identity. It should communicate uncertainty and make inferred memories easier to challenge.
Context boundaries and expiry
People behave differently at work, in class, with friends, and in private. A memory helpful in one setting may be inappropriate in another. Users need project-specific contexts, temporary sessions, and a simple way to say “do not use this here.” Some memories should expire automatically; others should remain only while the user chooses.
Where this can fail
A deletion control may be misleading if copies remain in logs, derived profiles, or connected systems. The interface should explain what deletion covers and what it does not. Memory systems should not quietly infer sensitive attributes or make them available to unrelated tasks.
Questions for research
- Which memory controls do people understand and actually use?
- How should a system communicate where a remembered detail came from?
- When should a memory expire automatically?
- Does inspectable memory improve trust or increase the feeling of surveillance?
Design hypothesis
Personalisation works best when memory is visible, scoped, correctable, and revocable.
Research starting points
- Amershi et al. (2019), CHI. https://doi.org/10.1145/3290605.3300233
- Google People + AI Guidebook. https://pair.withgoogle.com/guidebook/
- NIST Privacy Framework. https://www.nist.gov/privacy-framework