Study: ChatGPT memory mostly saves details without being asked, often about users' psychology

ChatGPT's memory saves information about users mainly on its own initiative rather than at their request. This is according to a study by researchers from Germany, India and the Netherlands, reported by the Polish tech site Spider's Web.

What the researchers found

The analysis used data voluntarily shared by 80 regular ChatGPT users, who downloaded their conversation history and passed it to the researchers. Materials from 65 participants contained 2,050 memory entries, created during 1,058 conversations.

Only 4.1% of entries were created after an explicit request from the user. The remaining roughly 96% were created by the system itself, which judged the information useful in future conversations. The results are described in the paper "The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT", prepared by researchers including some from the Max Planck Institute for Software Systems.

This does not mean ChatGPT broke a user's ban. Automatic remembering was an intended feature. The question is whether users understood how much could be saved without being told to.

Psychology and sensitive data

According to the study, 52% of entries contained information about users' psychological traits, such as desires, intentions, beliefs or emotions. A further 28% contained personal data as defined by GDPR. About 7% contained special categories of data, including health information. The authors called this collection an "algorithmic self-portrait".

Not every entry faithfully reflected what the user said. In one example, a person asked for bands similar to Nirvana, and ChatGPT saved that the user likes Nirvana. According to the researchers, 84% of entries were directly grounded in users' statements, 5% were judged weakly justified and 4% unsupported by context. A wrong assumption stored in memory can influence later answers.

The memory has already changed

The study covers an older mechanism, so its results cannot be applied fully to the current version. In June 2026, OpenAI introduced a system called "dreaming", which combines information from different conversations and updates it over time. According to Spider's Web, there is no comparable independent study showing whether the new memory stores sensitive data less often or draws wrong conclusions less often.

What this means for you

The memory feature is available to all users, including in the UK. According to Spider's Web, it can be switched off in ChatGPT settings, in the "Personalisation" section. In available versions of the service, you can also review a summary of remembered information and ask for it to be corrected.

Switching memory off does not automatically delete conversation history. To remove a particular piece of information completely, you may also need to delete chats, files or other places where you provided it. The recently introduced PIN lock for private conversations makes it harder for others to see them by accident, but does not delete data.

The study's authors expect the problem to grow, because chatbots are increasingly used as advisers on work, financial and personal matters.