Back
NoteApr ’242 min read

Memory needs a list

ChatGPT’s memory reached every Plus user this week, with a note when it saves something and a page to forget it. A product that remembers you owes you the list, in plain words.

Typeface

Yesterday OpenAI turned on memory for all ChatGPT Plus users, outside Europe and Korea, after testing it with a small group since February. ChatGPT can now carry things you’ve told it from one conversation to the next.

The design details are better than I expected. When it saves something, the chat shows “Memory updated.” You can open a list of what it has saved, forget items one at a time, or turn memory off. Temporary chats don’t use it. OpenAI says it will steer away from proactively remembering sensitive details like health information.

I think the list is the most important part of the launch, more important than the remembering.

Personalization has mostly been invisible. A feed learns what you like and never tells you what it decided. You find out sideways, when it keeps showing you something you looked at once. You can’t correct a belief you can’t see.

A product that remembers you owes you the list of what it remembers, in plain words, with a way to delete each line. ChatGPT’s list does that. Each memory is a sentence you can read, and each one can go.

My worry is timing. “Memory updated” appears after the save. To see what was written, you have to go and look.

On Uber’s experimentation platform, the rule I took from the rollback dialog was to show the consequence before the commitment. A memory is a small commitment, but it shapes every answer after it. I’d want the note to say what was saved, inline, with an undo right there, the way a good email client lets you take back a send.

The cost of that is noise. People will get tired of being told. That’s why the right default is probably quiet for small things and explicit for anything personal. Where that line sits is the design question, and I’d rather see OpenAI draw it in the product than in a policy.

There’s one thing the list doesn’t settle. The Register quotes OpenAI saying it may use content you provide, “including memories, to improve our models.” Deleting a line from your list and keeping it out of training are two different promises. People will assume one includes the other, so the product should say plainly whether it does.