# ChatGPT saved most memories without being asked, study finds Memory makes ChatGPT more useful, and this study shows how much of it happens without a request from the user. The Assurance Desk, news. Published 21 October 2026. ChatGPT created 96% of the memories it kept about users without being asked, a study of 2,050 memories has found. Researchers in Germany and the Netherlands first published it on 1 February 2026. The paper has been accepted at The ACM Web Conference 2026 ([arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (full paper)](https://arxiv.org/html/2602.01450v2) and [arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (abstract page)](https://arxiv.org/abs/2602.01450), checked 21 October 2026). ## ChatGPT made most memories on its own The researchers looked at 2,050 memory entries from 80 ChatGPT users. Only 4% of the memories, 84 in all, came from a user asking. ChatGPT started the other 96% ([arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (full paper)](https://arxiv.org/html/2602.01450v2) and [arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (abstract page)](https://arxiv.org/abs/2602.01450), checked 21 October 2026). They counted a memory as asked for when the user's message used words such as remember, save or forget ([arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (full paper)](https://arxiv.org/html/2602.01450v2), checked 21 October 2026). Most memories matched the conversation. 84% came straight from something the user had said ([arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (abstract page)](https://arxiv.org/abs/2602.01450), checked 21 October 2026). **What the 2,050 memories in the study held** - Created by ChatGPT, not asked for: 96% - Grounded in what the user said: 84% - Psychological insights: 52% - GDPR-defined personal data: 28% - Special category data: 7% Share of all 2,050 memory entries from 80 users. One memory can fall into more than one group, so the bars are not parts of one whole. Source: arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (abstract page), https://arxiv.org/abs/2602.01450 (checked 21 October 2026); arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (full paper), https://arxiv.org/html/2602.01450v2 (checked 21 October 2026) ## About half the memories were about the person 52% of the memories held psychological insights about the user. 28% held personal data as the GDPR defines it ([arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (abstract page)](https://arxiv.org/abs/2602.01450), checked 21 October 2026). 7% held special category data. 35% of the users had health information saved in their memory ([arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (full paper)](https://arxiv.org/html/2602.01450v2), checked 21 October 2026). The sample is small. The users came through Prolific and were mainly in the United States and Europe. The authors say their methods, which used AI models to label the data, are imperfect ([arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (full paper)](https://arxiv.org/html/2602.01450v2), checked 21 October 2026). ## OpenAI says memory can save useful context OpenAI's help page says saved memories are details you ask ChatGPT to remember, or details it saves as useful context ([OpenAI Help Center, Memory in ChatGPT](https://help.openai.com/en/articles/8590148-memory-in-chatgpt), checked 21 October 2026). It also says sensitive information can appear in memory if you share it with ChatGPT ([OpenAI Help Center, Memory in ChatGPT](https://help.openai.com/en/articles/8590148-memory-in-chatgpt), checked 21 October 2026). The paper describes OpenAI's policy as not storing sensitive data, such as health information, unless the user asks ([arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (full paper)](https://arxiv.org/html/2602.01450v2), checked 21 October 2026). ## For UK organisations, the usual rules apply The UK GDPR gives some personal data extra protection, and health data is one of the special categories ([ICO, What is special category data?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/what-is-special-category-data/), checked 21 October 2026). The ICO has answered questions about AI tools at work. Privacy notices should say how a tool is used. Outputs must be checked for accuracy. Special category data needs a second condition under Article 9 of the UK GDPR ([ICO, Previously asked questions](https://ico.org.uk/for-organisations/advice-and-services/innovation-advice/previously-asked-questions/), checked 21 October 2026). Workspace owners and admins can manage the memory settings their plan supports ([OpenAI Help Center, Memory in ChatGPT](https://help.openai.com/en/articles/8590148-memory-in-chatgpt), checked 21 October 2026). **Checking ChatGPT memory today** - Open Settings, then Personalization, then Memory, and read what is there. - Ask ChatGPT what it remembers about you. - Delete a saved memory itself. Deleting the chat alone may not remove it. - Use a temporary chat when you do not want memories made. Steps from OpenAI's help page. Controls vary by plan, region, platform and workspace. Source: OpenAI Help Center, Memory in ChatGPT, https://help.openai.com/en/articles/8590148-memory-in-chatgpt (checked 21 October 2026) ## What happens next The ICO's guidance on special category data is under review because of the Data (Use and Access) Act ([ICO, What is special category data?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/what-is-special-category-data/), checked 21 October 2026). The researchers also built a tool, Attribution Shield, that warns before a question reveals something sensitive and suggests a safer wording ([arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (abstract page)](https://arxiv.org/abs/2602.01450), checked 21 October 2026). ## Read next - [Data protection: the section](https://theassurancedesk.com/data-protection) - [Should staff use ChatGPT with memory switched on?](https://theassurancedesk.com/guides/chatgpt-memory-at-work) - [Can staff paste client data into ChatGPT or Copilot?](https://theassurancedesk.com/guides/staff-client-data-chatgpt-copilot) - [A DPIA for an AI tool: when and how](https://theassurancedesk.com/guides/dpia-for-an-ai-tool) - [What must your UK privacy notice say?](https://theassurancedesk.com/guides/what-your-privacy-notice-must-say) ## Sources - arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (full paper): https://arxiv.org/html/2602.01450v2 (checked 21 October 2026) - arXiv, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (abstract page): https://arxiv.org/abs/2602.01450 (checked 21 October 2026) - OpenAI Help Center, Memory in ChatGPT: https://help.openai.com/en/articles/8590148-memory-in-chatgpt (checked 21 October 2026) - ICO, What is special category data?: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/what-is-special-category-data/ (checked 21 October 2026) - ICO, Previously asked questions: https://ico.org.uk/for-organisations/advice-and-services/innovation-advice/previously-asked-questions/ (checked 21 October 2026) This is an independent editorial guide, not legal advice. --- **Who publishes this.** The Assurance Desk is published by Add-Venture Advisory Ltd, company number 11854524, registered in England and Wales. The same owner runs Avouch, a marketing agency that sells AI visibility services to firms of the kind this site covers. That is a conflict of interest, and it is disclosed here.