AI memory should feel like a shelf, not surveillance.
An assistant becomes more useful when it can remember a preference, deadline, recommendation, or unfinished thought. It also becomes more sensitive. The answer is not pretending memory has no risk; it is designing memory so the person can see and control it.
Try it with Cappy →
Chat history is not the same as memory
Chat history is a transcript. Saved memory is a selected set of facts or artifacts that may be retrieved in future conversations. Keeping the two separate makes behavior easier to understand: deleting a conversation should not secretly preserve every personal detail, and saving one useful fact should not require keeping an entire transcript forever.
Three reasonable memory modes
An explicit-only mode saves something only after a phrase such as ‘remember this.’ A suggestion mode prepares a candidate and asks before saving. An automatic mode can capture likely-useful details but should still expose them in an inbox or activity log. The safest default for a new product is explicit or suggestive memory.
Controls worth demanding
People should be able to browse memories, search and filter them, correct a summary, change categories, merge duplicates, delete one item, or erase everything. Retrieval should prefer relevant memories instead of injecting the entire personal archive into every prompt. An export should preserve the data in a portable form.
A useful test
Tell the assistant one temporary fact, one durable preference, and one thing it should not save. Then inspect its memory view. Ask the same topic later, edit the preference, and delete it. If the behavior is surprising or the controls are missing, the product is asking for more trust than it has earned.
Quick answers
What should an AI assistant remember?
Details with an obvious future use, such as a stable preference, project context, important date, or deliberately saved recommendation. Passing emotion and sensitive inferences should have a much higher bar.
What controls should memory provide?
At minimum: view, search, edit, delete, export, and a choice between explicit, suggested, or automatic saving. Useful systems also merge duplicates and show relevant retrieval.
Related guides
AI Chat Privacy Checklist: 12 QuestionsUse this practical checklist to evaluate an AI chat app's storage, model providers, memory, deletion, exports, search, and voice handling.
What Should an AI Companion Remember?Explore useful and inappropriate forms of AI companion memory, including preferences, plans, recommendations, artifacts, and sensitive details.How On-Device AI WorksLearn what on-device AI means, what stays local, its privacy benefits, its limits, and how browser-based local AI works.