How does ChatGPT's memory of past chats work?

Since February 2024, ChatGPT has been able to keep a short list of facts about you, such as "owns a coffee shop" or "prefers meeting notes as bullet points," and use them in later conversations. On April 10, 2025, OpenAI began rolling out a broader version. ChatGPT can now refer to the content of your previous conversations as well, so it can pick up your interests, your usual tone and the topics you return to without you repeating them.

The two parts work differently, and that difference matters when you want to know what ChatGPT is using.

Part one: saved memories

Saved memories are individual notes that ChatGPT writes down, either because you asked ("Remember that I teach grade 3") or because it judged something worth keeping. They are visible. You can open the list, read every entry and delete the ones you don't want. Deleting a conversation does not delete a memory that came from it; you have to remove the memory itself. When OpenAI introduced the feature, it said it steers ChatGPT away from saving sensitive details, such as health information, unless you ask it to.

Part two: reference chat history

The newer part has no list. In an email quoted by Decrypt, OpenAI explained that this part learns your tone, goals and interests from earlier chats, keeps evolving, and, unlike saved memories, is not displayed as a list in settings. In practice, ChatGPT can look back across your old chats when answering a new question, rather than relying only on the notes it saved.

According to TechCrunch, the feature also reaches ChatGPT's voice and image-generation features, not only text. It is rolling out first to paying subscribers; reports differ on whether Plus users get it at once or after Pro users. It is not yet offered in the UK, the European Economic Area, Switzerland, Norway, Iceland or Liechtenstein, where OpenAI says further reviews are needed. If you already had memory switched on, the new feature is on by default.

A useful way to picture it: saved memories are a notebook you can read; chat history is the assistant having skimmed your past conversations. You can switch either off, but only the notebook can be read page by page.

What changes in everyday use

The upside is less repetition. If you have spent months asking for help with your bakery's social media posts, a new request like "draft this week's post" may already come back in your usual style and length.

The downside is that old context can leak into new answers in ways you didn't intend. A conversation where you were testing a silly idea, or helping a friend with their résumé, can shape later replies as if it described you. OpenAI has not said how long chat history is kept for this purpose, according to Decrypt.

How do you check and manage what it uses?

Run a short review. It takes ten minutes and tells you far more than reading about the feature.

Step 1: ask directly. In a normal chat, send:

Tell me everything you know or assume about me from our past conversations and your saved memories. Group it under: work, family, location, interests, writing preferences, and anything else. For each item, say whether it comes from a saved memory or from earlier chats.

ChatGPT's answer is its own summary, so treat it as a guide, not a complete record. Look for anything wrong, out of date, or more personal than you are comfortable with.

Step 2: open the saved list. In the settings, find Personalization and then the memory options (the original path was Settings > Personalization > Memory > Manage). Delete any saved memory that is wrong or that you don't want used. Remember that deleting the chat it came from is not enough.

Step 3: decide on chat history. In the same settings area you can switch off the new past-chats feature and manage individual saved memories, according to TechCrunch, or turn memory off entirely.

Step 4: correct in plain words. For a single wrong assumption, you can also tell ChatGPT: "Forget that I live in Toronto; I moved to Montreal." Then repeat Step 1 a few days later to see whether the correction took.

Step 5: use a temporary chat for one-offs. A temporary chat neither draws on memory nor adds to it, and does not appear in your history. Use it for anything you don't want shaping future answers, such as a friend's résumé or a sensitive question. One detail from OpenAI's earlier announcement still applies: custom instructions, if you have set them, are followed even in a temporary chat.

The same review is worth doing before you use one account for both work and personal questions, since anything ChatGPT has picked up in one area can shape answers in the other.