Overview
How might we let people choose how far a fact persists before it becomes durable assistant context?
When to use
- Essential for AI chat products with memory features, especially in enterprise and team contexts where retention policies vary by task and sensitivity.
When to skip
- Stateless products with no memory layer at all.
- Single-session throwaway chats where every option is “this chat only.”
- Enterprise modes where policy already forces chat-only and a toggle would lie.
States
Design the scope decision at save time, not as a buried settings toggle after the fact is already stored.
- 01
Proposed
The assistant or user identifies a fact worth remembering. Nothing is stored yet.
- 02
Scoping
Save to options appear: this chat, project, or all chats. The person picks a destination.
- 03
Confirming
A short line confirms where the fact will live and what that scope means in plain language.
- 04
Saved
Explicit save commits the fact at the chosen scope. The UI reflects the destination.
- 05
Reviewing
Settings or in-thread review shows what is already remembered at each scope.
- 06
Revoked
The person removes or narrows scope. Downgrade from all chats to project should be one clear action.
Key UX elements
The parts that must be present for scope choice to prevent accidental long-term memory.
Fact
Show the exact text that will be stored.
Quote the preference or detail verbatim so people know what they are promoting into memory.
Scope
Offer distinct Save to destinations.
This chat, project, and all chats must read differently. Avoid labels like Saved vs Remembered.
Confirm
State where it will live before commit.
Update confirmation copy as the scope changes so promotion never feels silent.
Save
Make storage explicit.
A primary Save action beats auto-remember on send, especially for pasted or sensitive content.
Review
Let people inspect stored memory.
After save, link to a review list grouped by scope so users can edit or delete what the assistant knows.
Default
Default to the narrowest safe scope.
Chat-only or project-only defaults reduce risk when someone pastes credentials or one-off context.
Rules
Defaulting sensitive pastes into long-term memory without a scope choice.
Toggles that do not show what is already remembered after send.
Silent promotion from project memory to org-wide memory.
Scope labels users cannot distinguish (Saved vs Remembered vs Learned).
Evidence
| Product | Implementation |
|---|---|
| Claude Projects | Project-scoped knowledge separate from global chat memory. |
| ChatGPT Memory | Controls for what is remembered and ability to manage saved memories. |
| Notion AI workspaces | Workspace and page context boundaries for AI assistance. |
| Enterprise copilots | Tenant and channel scopes with admin-enforced retention. |
