Will AI Assistants Know How to Forget, Too?

At first, it’s convenient. The assistant remembers which software we use, that we prefer short answers, and that we’re preparing a presentation. Finally, we can ask for something without introducing ourselves all over again. Then we change jobs, the project wraps up, a worry fades. And that memory, so useful until yesterday, starts to feel uncomfortable.

With another person, we can say, “Forget about it. That’s ancient history.” With an AI assistant, the same sentence raises a technical question: which information should disappear? The detail in the conversation, the one saved in our profile, or the one still retrievable from a connected document? A simple “forget” can involve very different operations.

So there’s already a first answer to the question in the headline: some tools let us delete memories and correct personal information. What remains unclear is how complete that forgetting really is, and whether assistants will also learn to recognise when a detail has stopped being useful.

Memory Isn’t Just One Drawer

Talking about memory as a single thing is convenient, but potentially misleading. There’s the text an assistant is using in the current conversation. There’s information retained to personalise future interactions. And there are the archives it can search: old chats, files, connected applications.

Then there’s what the model learned during training. This distinction matters: deleting a preference from your profile and removing a piece of information’s influence from a neural network’s parameters are different problems.

To understand what happens when we press “delete”, we therefore need to know where the information was stored. The interface may have a single button; the underlying system may have several places to look.

Deleting a Chat May Not Be Enough

ChatGPT’s documentation provides a concrete example: deleting a conversation does not necessarily erase a memory saved separately. Removing information may also require deleting saved memories and addressing other sources that contain it. Turning memory off, meanwhile, does not delete chat history.

This is where product language can create the wrong expectations. “Stop using this detail”, “don’t mention it” and “delete it” sound like similar requests. The first concerns the assistant’s behaviour; the last concerns data retention. A system can stop mentioning something without deleting its source.

Timing matters, too. OpenAI notes that updates and deletions can take time to propagate, and that logs of deleted saved memories may be retained for up to thirty days for safety and debugging. A chatbot’s silence, on its own, does not prove that every copy has disappeared.

An Unhelpful Memory Can Still Be True

A memory doesn’t have to be inaccurate to become uncomfortable. Sometimes, the information has simply stayed the same while the person has changed.

Imagine an assistant that keeps suggesting budget options because, months earlier, we mentioned going through a difficult financial period. Or one that interprets every professional request through the lens of a job we left long ago. The original detail was true; its later use has become misplaced.

A good assistant should distinguish a lasting preference from a temporary need. “I’m vegetarian” and “I’d like a meat-free dinner tonight” don’t justify the same conclusion. It would help to give memories an expiry date, restrict them to a project, or require confirmation before reusing sensitive details. These are design principles before they are new features to advertise.

Making a Model Unlearn Is a Different Matter

When the aim is to change what a model learned during training, we enter the field of machine unlearning. Researchers are studying how to remove the influence of particular information without repeating the entire training process from scratch.

The difficulty is that a model doesn’t store everything as a collection of removable index cards. Changing its behaviour and verifying that information has actually been removed require different kinds of evidence. Blocking an explicit answer does not, by itself, guarantee that the content cannot resurface through other questions.

A paper by Google DeepMind researchers highlights precisely this gap between expectations of unlearning and what the techniques can deliver. Presenting it as a universal eraser would promise too much.

An Assistant Should Let Us Change

Memory makes a service more convenient, but it also gives that service the power to decide which parts of the past to bring into the present. That calls for clear controls: we need to see what is remembered, understand where it came from, correct it, and know what deleting it will do.

The most interesting shift will be from simply accumulating information to choosing what deserves to stay. An assistant should know when to ask, “Does this still apply?” before treating an old confidence as a permanent characteristic.

Remembering the name of our project saves time. Recognising that the project is over takes something more. Part of AI assistants’ maturity will be measured by their ability to let information go: how much room they leave for who we have become.