Overview
How might we let people inspect, edit, and delete what the assistant remembers across sessions without memory feeling like hidden surveillance?
When to use
- Personal assistants like ChatGPT saved memories or Claude memory where facts carry forward between chats.
- Products where users need to fix stale preferences, wrong inferences, or outdated context the model keeps applying.
- Long-term agent relationships where transparency about stored facts builds trust more than silent continuity.
- Any product that stores preferences or inferred facts and needs a single place to review them.
When to skip
- Stateless tools where remembering prior chats adds risk without value.
- Regulated contexts that forbid durable personal memory without enterprise controls.
- Products that only need short thread context, not cross-session profile memory.
States
Design the memory review surface, not only the moment something gets saved.
- 01
Listed
Saved memories appear in a dedicated settings or profile view. Count, search, and sort make the list scannable.
- 02
Inspecting
The person opens one memory to read the exact text, category, and when it was last updated or used.
- 03
Editing
Inline edit fixes wording before it influences future replies. Save commits the revised fact.
- 04
Adding
Manual add lets people declare a preference or fact the assistant missed. New entries join the same list.
- 05
Deleted
Remove confirms before the fact disappears. The assistant should stop applying it on the next turn.
- 06
Paused
Memory is off or limited. Copy explains pause versus wipe so people know whether stored facts remain.
Key UX elements
The parts that must be present for stored memory to feel controllable.
List
Show every stored fact in one place.
A scrollable list with count and search beats hunting through chat history for what the assistant knows.
Edit
Fix wording inline.
Let people correct a memory’s text before it keeps shaping answers. Save should feel immediate.
Delete
Remove with confirmation.
One clear delete action plus a short confirm. Do not bury removal three levels deep in settings.
Add
Allow manual memory entry.
Some facts never get inferred. A simple add field covers preferences the user wants declared up front.
Category
Separate preferences from inferred facts.
Label type or source so people can tell declared context from model guesses.
Pause
Explain off versus wiped.
When memory is disabled, say whether facts are frozen, hidden, or permanently deleted.
Rules
Invisible memory that changes answers with no settings entry.
No way to delete or export individual memories.
Mixing declared preferences with inferred facts in one unlabeled blob.
Turning memory off without explaining pause vs wipe.
Evidence
| Product | Implementation |
|---|---|
| ChatGPT | Saved memories list with search, prioritize/delete, and version restore. |
| Claude | Opt-in memory under Capabilities with import and pause-vs-reset. |
| Perplexity | Search-history memory toggle with Max-tier Brain upsell. |
| Gemini | Activity and personalization controls tied to Google account settings. |
