How does ChatGPT's new memory work, and how do you control it?

Until now, every new ChatGPT conversation started from zero. If you told it on Monday that you teach kindergarten, it had no idea on Tuesday. Yesterday, February 13, 2024, OpenAI announced that it is testing memory, a feature that lets ChatGPT carry useful details from one conversation to the next.

The test is small at first. This week OpenAI is switching it on for a small share of people on the free and Plus plans, to see whether it helps, and says it will share plans for a wider rollout later. If you do not see it yet, that is expected.

What memory does

With memory, ChatGPT can keep short notes about you and apply them later. You can ask it directly ("Remember that I prefer metric units"), or it can pick up details on its own as you chat. OpenAI says the memory improves the more you use ChatGPT.

OpenAI's own examples show the idea. If you have explained that you like meeting notes with headlines, bullet points and action items at the end, ChatGPT can format future recaps that way. If you mentioned that your toddler loves jellyfish, it can suggest a jellyfish in a party hat when you ask for help with her birthday card. A teacher who prefers 50-minute lessons with follow-up activities would not need to repeat that for every lesson plan.

Here is how memory compares with the tools ChatGPT already had:

Feature	What it stores	Who fills it in	Best for
Chat history	Each past conversation, separately	Saved automatically	Going back to an old conversation
Custom instructions	A fixed profile and response style	You, in settings	Standing rules: "keep answers short"
Memory (new)	Short facts gathered across all chats	You or ChatGPT, during chats	Details that come up naturally over time
Temporary Chat (new)	Nothing kept in history or memory		Private or one-off questions

How you stay in control

OpenAI stresses that the user is in charge, and gives several ways to manage memory:

Tell it. Say "remember this" or "forget that" in plain language, or ask "What do you remember about me?"
Review it. In Settings > Personalization > Manage Memory, you can see individual memories, delete any one of them or clear them all.
Turn it off. In Settings > Personalization > Memory, you can switch the feature off. While it is off, ChatGPT neither creates nor uses memories.
Use Temporary Chat. A new kind of conversation that does not appear in your history, does not use memory and is not used to train OpenAI's models.

One detail matters: memories are not tied to the conversation they came from. Deleting a chat does not delete what ChatGPT learned in it; you have to remove the memory itself. TechCrunch also points out that, for people in the test, memory is switched on by default, and that Temporary Chat still follows your custom instructions if you have set them. For safety reasons, OpenAI may hold on to Temporary Chat conversations for as long as 30 days.

Privacy and training

Memory raises obvious questions: what should a chatbot remember, and who sees it? OpenAI says it is training ChatGPT not to remember sensitive information, such as health details, on its own initiative, unless you explicitly ask it to.

Content you share, including memories, may be used to improve OpenAI's models, unless you turn that off in your Data Controls. For ChatGPT Team and Enterprise customers, OpenAI says it does not train on their content at all, and an Enterprise account owner can disable memory for the whole organization. Team and Enterprise users will receive memory as part of the wider rollout.

GPTs, the custom chatbots found in the GPT Store, will get their own separate memories, which builders can choose to enable. A GPT's memory is not shared with ChatGPT or with the person who built the GPT. That also means a GPT will not know what you told ChatGPT; you may need to repeat relevant details.

Why it matters

Memory moves ChatGPT a step from a tool you brief each time toward an assistant that knows your context. For people who use it daily, that saves typing and gives more relevant answers from the first message.

It also makes privacy habits more important. A chatbot that remembers can accumulate a detailed picture of you over months. The controls OpenAI describes, viewing, deleting, switching off and temporary chats, are the right ones to learn now, while the feature is still new and the number of memories is small.

Try it: If memory appears in your settings, tell ChatGPT one harmless preference, such as "Remember that I like examples from cooking." Start a new chat, ask it to explain a math idea and see whether the example changes. Then open Manage Memory and delete that entry.

What should you let ChatGPT remember, and what should you keep out?

Good candidates are preferences and context that make answers better and would not hurt you if someone else saw them: your preferred format and tone, your profession in general terms, the programming language you use, the age group you teach, or the fact that you prefer metric units.

Keep out anything you would not write on a postcard: passwords, account and card numbers, government ID numbers, detailed health information, and private details about other people, especially children. Even with OpenAI's efforts to avoid storing sensitive details, the safest memory is one that never contains them.

Make reviewing memory a habit. Every few weeks, open Manage Memory, delete what is outdated or too personal, and use Temporary Chat for anything you would rather not have remembered at all.