PTN—38
Legible memory
Show people exactly what the system remembers about them, in plain language they can read, edit, and delete.
When an AI remembers things, the memory itself becomes part of the interface. If people can’t see what’s stored, every personalized response is a small jump scare: the assistant knows your kid’s name, your codebase conventions, that project you abandoned — and you have no idea what else. Legible memory turns the store into a first-class surface: a readable list of remembered facts, each one editable and deletable, written the way a person would say it.
The bar is “plain language,” not “data export.” A JSON dump of embeddings satisfies a compliance checkbox and nothing else. ChatGPT’s memory panel got this right: discrete sentences like “Prefers concise answers” that you can delete individually. The moment memory is written for humans, it also becomes correctable — people fix wrong memories instead of fighting their downstream effects forever.
Legibility also disciplines the system. If a remembered fact would look creepy written down — inferred income bracket, guessed health condition — that’s a signal you shouldn’t be storing it, or at minimum shouldn’t be acting on it silently. The memory panel is where your data practices meet daylight; design it as if a journalist will screenshot it, because eventually one will.
The pattern emerged the moment assistants crossed from stateless chat to persistent memory — ChatGPT’s memory launch made the store a product surface overnight, and every team since has faced the same tradeoff: memory aggressive enough to be useful is aggressive enough to be creepy, and legibility is the only lever that moves both sides at once. Maturity shows in the write path, not the read path. Immature products let you view memories; mature ones announce each write inline, label inference apart from instruction, and prove that deletion propagates. When people start deliberately teaching the memory — “remember that we ship Fridays” — the surface has earned its trust.
When to use
- The product persists any user-specific facts across sessions.
- Personalized behavior could plausibly surprise someone (“how did it know that?”).
- People need to correct stale facts — a changed job, a finished project, a moved city.
- Regulatory or trust stakes make “what do you know about me?” a question you must answer well.
Design considerations
- 01Write each memory as a single human-readable sentence, not a key–value pair.
- 02Let people delete any single memory in one action, and everything in two.
- 03Announce when a memory is saved, inline and at the moment it happens — “Memory updated” beats a silent write.
- 04Make memories editable, not just deletable; correcting “uses React” to “uses Svelte” is cheaper than delete-and-reteach.
- 05Distinguish explicit memories (“remember that I…”) from inferred ones, and label the inference.
- 06Let people say “don’t remember this” in the conversation itself, not only in settings.
- 07Show which memories influenced a given response when someone asks why the answer looks personalized.
Pitfalls
- ✕Memory that only surfaces in settings, three levels deep — people discover it exists when it misfires, which is the worst possible introduction.
- ✕Storing inferences as facts. “Interested in crypto” from one question becomes a permanent lens on every answer.
- ✕Delete that doesn’t propagate — the memory disappears from the panel but the behavior persists because it was baked into a profile embedding.
- ✕Remembering things people said in frustration or during a one-off task, then cheerfully replaying them weeks later.
In the wild
- ChatGPT
- Maintains a memory panel of plain-language facts; shows “Memory updated” toasts inline when it saves something, and lets users delete individual memories or wipe all of them.
- Claude
- Project instructions and preferences are user-authored text the person can read and edit verbatim — memory as a document you own rather than a profile inferred about you.
- Notion AI
- Answers questions grounded in workspace content the user can already see and edit, so “what it knows” maps directly to pages rather than a hidden profile.
- Google Gemini
- Offers a saved-info surface where users can view and remove details Gemini uses to personalize responses.