Study: 96% of ChatGPT memories were saved unasked

Short answer: A peer-reviewed study of real ChatGPT data found that 96% of the memories ChatGPT saved about its users were created without the user asking it to remember anything. About half of those memories described people's emotions, beliefs, desires or intentions, and more than a quarter contained personal data as defined by EU privacy law. The study looked at an older version of ChatGPT's memory, and OpenAI has since made memory more automatic, not less. You can review, correct and delete what it keeps, but deleting a chat does not necessarily delete what was remembered from it.

What happened

Researchers from the Max Planck Institute for Software Systems, the Max Planck Institute for Security and Privacy and several other institutions published a paper called The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT. It appears in the Proceedings of the ACM Web Conference 2026. This week the findings reached a wider audience through Gizmodo (October 22, 2026) and Futurism (October 25, 2026).

The team worked with ChatGPT histories donated by 80 real users, who obtained copies of their own data using their rights under European privacy law (GDPR). In the histories of 65 of them, the researchers found 2,050 memory entries, the notes ChatGPT keeps about a person so it can personalize later answers.

What the researchers found

• Most memories were not requested. Only 4% of memory entries (84 of 2,050) came after a clear request such as "remember this" or "save that". The other 96% were created or updated by ChatGPT on its own.
• Memories are personal. About 28% contained personal data as defined by the GDPR, and 52% contained what the authors call psychological insights: a person's emotions, beliefs, desires or intentions.
• Most were accurate, some were guesses. 84% were directly grounded in what users had said. Some were assumptions. When one user asked for "bands like Nirvana", ChatGPT saved that the user likes Nirvana, something the person never actually said (the system did mark it as weakly supported).

The authors call the result an "algorithmic self-portrait": a picture of you assembled from things you told a chatbot in what felt like private conversations.

Is this still how ChatGPT works?

Not exactly, and that matters. The study examined an earlier memory system. Both Gizmodo and Futurism point out that OpenAI has since introduced an upgraded system that continually synthesizes information from past chats instead of keeping only a list of individual notes.

OpenAI's own help page now describes two experiences. "Improved memory" keeps updating a broader summary of relevant context, while "legacy saved memories" use an explicit list of items. Depending on your plan and region, memory can draw on past chats, saved memories, custom instructions, files in your Library and connected apps such as Gmail. OpenAI also says plainly that the memory summary you can view "does not necessarily include every detail or source that Memory can reference."

So the study's core point still stands: much of what a chatbot remembers about you is decided by the system, not by you.

Why it matters for people who talk to AI about their lives

Many people use chatbots less like a search engine and more like a friend, a coach or a diary. Lead author Abhisek Dash told Gizmodo that the system can end up "acting as your advisor, your friend, your doctor, your lawyer, and so it has information on everything."

Lorrie Cranor, director of Carnegie Mellon University's CyLab Security and Privacy Institute, compared it to telling secrets to a stranger on a train. The difference, she said, is that the stranger is unlikely to be recording the conversation, while "the AI definitely is recording the conversation."

There is also a commercial side. Gizmodo reports that OpenAI began introducing ads in ChatGPT in February and says it can use past conversations and memories to personalize ads for users who have personalized advertising turned on. OpenAI says advertisers do not get access to conversations or memories and that ads do not influence ChatGPT's answers.

What to do today: a quick checklist

1. Look at what it remembers. In ChatGPT, open Settings, then Personalization, then Memory. You can also simply ask ChatGPT what it remembers about you.
2. Correct or delete. Fix wrong entries and delete the ones you would rather it forget. "Don't mention this again" only changes what ChatGPT brings up. It does not delete the original source.
3. Delete in more than one place. OpenAI says deleting a chat does not necessarily delete a saved memory created from it. To remove something fully, delete the saved memory and the chat where you shared it, plus any file or connected app that contains it.
4. Use Temporary Chat for sensitive topics. Temporary chats do not create or update memories. Choose "Unpersonalized" before you start if you also do not want existing memories used. Saving a temporary chat turns it into a regular one.
5. Check model training separately. Memory settings and the "Improve the model for everyone" setting (in Data controls) are different switches. Turning one off does not turn off the other.
6. Rephrase personal questions. The researchers suggest asking about a situation rather than about yourself: "What should a person with diabetes do about X?" instead of "I'm diabetic, what should I do about X?"
7. Expect some delay. OpenAI says deletions can take time to propagate, and it may keep logs of deleted saved memories for up to 30 days for safety and debugging.

Frequently asked questions

Does ChatGPT remember everything I say? No. OpenAI says memory does not retain every detail and that ChatGPT decides what seems relevant. The study shows those decisions are mostly made by the system, not by explicit user requests.

If I turn memory off, are my chats deleted? No. Turning memory off does not delete past chats. If you turn it back on later, ChatGPT may create new memories from chats still in your history.

Can I see everything ChatGPT knows about me? Not fully through the summary. OpenAI says the summary is a high-level view that can leave out details. Under GDPR, people in Europe can request a copy of their data, which is how the study's participants got theirs.

Is this only a ChatGPT issue? The study only analyzed ChatGPT, because it was the first major chatbot to add memory. Other assistants and companion apps also remember things about users, so it is worth checking the memory and privacy settings of any app you confide in.

This article reflects the published study, OpenAI's help center and news coverage as of October 26, 2026. Memory features differ by plan, region and app version.

Related reading:

• Why does AI forget what you told it?
• Is ChatGPT safe? What happens to what you type
• ChatGPT Locked Chats: what the new PIN lock hides, and what it doesn't
• How to delete or export your AI chat history

Cover photo: Jernej Furman, "Hand holding smartphone with ChatGPT", via Wikimedia Commons / Flickr, CC BY 2.0.