Selective Memory
What is Selective Memory?
Selective Memory gives users explicit control over what AI remembers, forgets, or ignores. Instead of opaque memory, the system provides transparent controls to view, edit, or delete stored information. It's essential for personal assistants, conversational systems, or tools building context over time. Examples include ChatGPT's memory settings for viewing and deleting memories, Google Assistant's activity controls, or Replika marking conversations as temporary.
Problem
AI systems remember information without user visibility or control, risking privacy issues and inappropriate responses based on outdated or sensitive context.
Solution
Provide transparent memory controls letting users view, categorize (important/temporary/forget), and understand how stored information influences AI responses.
Real-World Selective Memory Examples
Implementation
When to use Selective Memory, and when it backfires
Use it when
- The product accumulates personal context over time (assistants, companions, long-running chats) and stale or wrong facts will degrade answers. Memory needs an off switch the user can reach.
- What the AI retains is sensitive or identity-shaping: health, relationships, finances, past mistakes. The user has a legitimate need to see it and revise it.
- Memory is the feature, not a side effect. If personalization is your pitch, you owe the user a window into what you kept and a lever to change it.
Don't, or minimize, when
- Nothing persists between sessions. Bolting a 'Memory' panel onto a stateless tool invents a control for data that does not exist and implies retention you are not doing.
- The only thing you can offer is hide-from-view. If delete cannot reach the embeddings, the fine-tune, the logs, and the cache, do not ship a button that says 'Forget.' Say 'Hide' and mean it.
- You are hoarding everything and calling the dashboard 'control.' Remembering every trivia detail while missing the two preferences that matter is not selective memory, it is a junk drawer with a search bar.
The trap
The ghost memory: a 'Forget' or 'Delete' that only drops the item from the list while the fact still lives in the model's store, the retrieval index, or the system prompt, and keeps shaping every answer. The user watches the card slide away, believes the thing is gone, then gets a reply that could only come from what they just 'deleted.' This is worse than no forget button at all. A product with no delete is merely opaque, the user knows the AI remembers. A fake delete is a lie the user acted on: they disclosed the secret, hit Forget, trusted it, and you kept it. The betrayal is not that you remembered, it is that you told them you didn't.
Take it into your own product
- 1
If 'forget' doesn't delete everywhere, don't call it forget.
A delete that drops the card but leaves the fact in the embeddings, the cache, or the system prompt is the worst outcome in this pattern: the user disclosed something, trusted the button, and you kept it. No delete is merely opaque. A fake delete is a betrayal the user acted on. Either purge every copy or label it 'Hide' and say so.
- 2
Show what you actually kept, including what you inferred.
Selective memory starts with a window, not a switch. List stored facts in plain language, when each was captured, and whether the user added it or the system guessed it. Silent inference is the thing users find creepy. If the AI decided to remember it, it belongs on the screen.
- 3
Remember what matters, not everything.
Hoarding every trivia detail while missing the two preferences that actually change answers is not control, it's a junk drawer with a search bar. Treat memory as a budget. Rank by what steers future output: stated preferences, constraints, corrections. Let the trivia expire.
- 4
Make memory's influence visible at the moment it acts.
A settings page buried three menus deep is not control. When an answer leans on a stored memory, show which one, right there, with a one-click 'this is wrong, forget it.' Control the user can reach mid-conversation beats a dashboard they have to go hunting for.
- 5
Forgetting is a feature, not a failure.
Teams treat deletion as data loss to be discouraged with friction and 'are you sure' walls. In a memory product the opposite is true: a clean, trusted forget is what makes users comfortable telling the AI anything in the first place. The off switch is what sells the on switch.
Add Selective Memory to your product
Copy the prompt below into Claude Code or Cursor in your repo. It encodes the four moves on the left and asks Claude to find your AI decision surfaces and update them. Claude reports what it changed and asks before adding dependencies.
Check if your product already has this pattern
Upload a screenshot. We'll tell you which of the 36 patterns your AI interface uses and where the gaps are.
Audit My Design
