Does ChatGPT Remember Everything You Tell It?

Original studyProceedings of the ACM Web Conference 2026 (WWW '26), 2026Read the paper →

Does ChatGPT remember everything you tell it? Not everything, but a new study suggests it decides for itself what to keep. Researchers analysed 2,050 memories from 80 real ChatGPT users and found that 96% were saved by the chatbot without being asked, and more than half held details about people’s desires, intentions, emotions, or beliefs.

Short answer: ChatGPT does not store every message, but its memory feature saves selected details from chats, mostly on its own initiative. In this study, only 4% of saved memories came from a user asking it to remember something. About half described psychological states, and most memories accurately reflected what users had actually said.

What the study looked at

Researchers at the Max Planck Institute for Software Systems, Ruhr University Bochum, the Max Planck Institute for Security and Privacy, the Indian Institute of Technology Kharagpur, and Delft University of Technology studied data donated by 80 Prolific participants, mainly in the US and Europe. Using their GDPR right of access, participants shared their ChatGPT histories. The team examined 1,058 conversations containing 22,971 exchanges, of which 2,050 triggered a memory update. The paper appears in the proceedings of the ACM Web Conference 2026.

Does ChatGPT remember everything you tell it without asking?

Only 84 memories, or 4%, followed an explicit request such as “remember” or “save”. The other 96% were created by ChatGPT itself. The researchers argue this shifts control over memory away from users.

Some 28% of memories contained personal data as defined under the GDPR, such as names, and 7% held special-category data. Health was the most common sensitive type, and 35% of participants had health-related information saved, despite OpenAI’s policy that such data should not be stored unless a user asks.

Your desires and intentions, saved

Using a psychological framework known as theory of mind, the team found that 52% of memories captured mental states. Desires were the most common (73% of these), followed by intentions (16%), emotions (14%), and beliefs (11%). Nearly all participants (97%) had at least one desire recorded. The authors note that desires and intentions are more lasting than passing emotions, so memories may build a durable picture of a person.

For anyone who talks to a chatbot about their mental health, this raises questions about self-disclosure and consent. A worry mentioned in passing could become a stored fact about you. Earlier work has explored why AI chatbots can help or harm mental health, and the psychological concerns people have about AI.

What the viral posts get wrong

Posts on X have suggested ChatGPT secretly profiles every user. The study does not show that. It covers 80 users, and 84% of memories directly matched users’ own words, with most of the rest paraphrased or reasonably inferred. Users can also review and delete memories. The authors acknowledge their sample is small and may not reflect all users, and that OpenAI’s system may change.

They also built a tool, Attribution Shield, that warns users about sensitive memories and suggests rewording queries; over 94% of reworded queries prevented attribution.