When AI remembers, your product decides what to forget

For two years the joke was that AI has no memory. You'd re-paste your taste file every session and hope it stuck. This week that excuse ended.

I've been writing that your AI agent has no memory and context files are the fix — you carried your own continuity. Now Anthropic's release notes for Claude memory that now works across chat and Cowork and TechCrunch's coverage of Claude Cowork remembering what you told it in chat show the shift: topics you teach in one surface show up in the other. Useful, and a new liability. When remembering is free, what you forget is the product decision.

Why does memory change the product job?

Memory turns a chat into a relationship, and relationships need boundaries.

Before this week, persistence was your job. You wrote an AGENTS.md, pasted a rubric, re-taught the model each time. Tedious, but safe — nothing stuck unless you put it there. Anthropic flipped the default. Claude now keeps memory across Chat and Cowork under Topics in Settings > Memory, and sensitive topics like health or beliefs stay out unless you opt in. For Free, Pro, and Max it is on. For Team and Enterprise it is off.

If you ship an AI feature that remembers, you are now a steward of what persists and who can delete it. Users will assume you kept the right things and blame you when you kept the wrong ones. Repeat customers stop re-explaining themselves, but the same convenience can keep a pasted API key you never meant to save.

What should your product actually remember?

Only the things you would want recalled without asking.

My filter: if recalling it would make the next output better without the user noticing, keep it. If it is transient, do not.

Remember taste and system. Type pairing, spacing, button set, voice, and the bar for good. That is what you already write in a taste file — now the agent can read it without being reminded.

Remember intent that repeats. The goal, not the words. "Prefers short, plain English for investor updates" is durable. "Use this sentence Tuesday" is not.

Remember corrections once. If the user fixed the same point twice — "we never use thin icons" — save the rule, not the transcript. Everything else stays in the session. A pasted CSV or debug log is useful for ten minutes, noisy forever. The Anthropic guide to building reliable evals for agents frames it well: score the output against what you decided was good, not how much context you carried. Memory that does not improve that score is weight.

How do you design memory users can trust?

Make it visible, editable, and scoped.

Visible means users see what you kept in one click. Anthropic lists every memory as a Topic you can open, edit, or delete in Settings > Memory. Copy that pattern — a "What I remember about you" page beats a hidden embedding.

Editable means delete is instant and complete across surfaces. Claude syncs memory between chat and Cowork, so your product needs the same promise.

Scoped means memory does not leak. Team and Enterprise memory is off by default because shared spaces need consent. Even for solo users, scope it to the project. One global memory that drags a freelance client's voice into another client's draft is a trust bug.

Add one UI line: "Memory changed this answer." When a recalled preference shaped the output, say so. Users forgive personalization they can see. They resent what they have to guess at.

When should your product deliberately forget?

When being wrong costs more than asking again.

Most products get this backwards. They remember everything and add a "clear history" button. That button is never pressed until after the awkward recall already happened.

Ship the opposite default: forget by default, remember by request. Pasted secret? Forget when the session ends. "Remember I prefer US English"? Keep it. Not sure? Ask — "Want me to remember this?" — and store the yes as a topic, not the chat.

Run a small pruning loop. Once a week, surface the three oldest memories and ask: still true? I described that curation habit in why curation is the job now — generation got cheap, keeping got expensive. Memory follows the same curve.

Do not build a clever memory system. Build one that is easy to inspect. One place to see it, one tap to delete it, one rule for what earns a spot. The models finally remember. Your job is to make sure they remember the right things.