ChatGPT Memory: Over 95% of Saved Entries Were Never Requested by the User

Written by

in

8 min read

A team from the Max Planck Institute for Software Systems, the Max Planck Institute for Security and Privacy, Ruhr University Bochum and Delft University of Technology read through 2,050 ChatGPT memory entries donated by 80 users. Only 84 of them, 4.1%, were saved because the user said something like “remember this.” ChatGPT wrote the other 96% on its own[1].

Seven months later a team at Rutgers University counted the same thing in a completely different population: 1,057 users in India, Nigeria, Brazil and Pakistan, 7,051 memory entries. The split was 95.36% system-initiated against 4.64% user-requested[2]. The two samples share no users, no countries and no recruitment method.

Memory is the part that survives after you close the chat window. Working memory inside a single conversation is the context window, which disappears when the conversation ends. Memory is stored separately and carries into the chats you start next week.

Memories are written one line at a time, mid-conversation

Both studies were possible because of the right of access under Europe’s General Data Protection Regulation (GDPR), which lets anyone demand a copy of the data a company holds and processes about them. Participants exercised that right, pulled their full ChatGPT histories, and donated them for research[1].

Inside ChatGPT the tool that writes memories is called bio. Every time it fires, the conversation log keeps a line reading model set context updated. The researchers found every one of those lines, then walked back up the thread to the user message that triggered the save[1]. That yielded 2,050 entries, drawn from 1,058 conversations belonging to 65 of the 80 participants. Those conversations contained 22,971 prompts, so within the chats that produced memories, roughly one prompt in eleven left an entry behind[1]. None of those saves involved a button press or a confirmation.

The system started almost every save

Share of memory saves started by the system in two studies
Share of memory saves started by the system in two studies

In the Max Planck set, 84 entries came from an explicit user request, and only 25 of the 65 participants had ever made such a request[1]. In the Rutgers set, user requests accounted for 4.64% of saves[2].

The decision about what gets kept has effectively left the user’s hands. The Max Planck authors describe user influence as mediated by what people happen to say in conversation, with no direct control over save events[1].

Names, financial circumstances, and notes on your state of mind

What the 2,050 memory entries contained
What the 2,050 memory entries contained

The Max Planck team classified entries against two GDPR definitions: personal data under Article 4(1), which covers names, identifiers, location and economic, social or cultural identity, and special category data under Article 9(1), which covers racial or ethnic origin, political opinions, religious beliefs, trade union membership, genetic and biometric data, health, and sex life or sexual orientation. The second tier is the one European law bars from processing unless a specific exception applies.

Across all 2,050 entries, 28% held personal data and 7% held special category data. Counted per person, 91% of participants had personal data sitting in memory, 54% had special category data, and 35% had health information stored[1]. Among the entries that held personal data, names showed up in 41% and economic, social or cultural identity in 40%. Health accounted for 61% of the special category entries[1].

The researchers also counted material that legal categories do not cover. They found descriptions of the user’s state of mind in 52% of entries and sorted them with theory of mind categories, a term psychologists use for working out what another person wants, believes or intends. Among entries with that content, desires and goals appeared in 73%, intentions in 16%, emotions in 14% and beliefs in 11%[1].

The paper quotes real entries. “User has been feeling lonely at times, despite having friends and family, and is seeking more companionship.” “User relies heavily on advice and recommendations from ChatGPT due to limited immediate access to their doctors and financial constraints.”[1]

The team also checked whether memory distorts what people said. It mostly does not: 84% of entries were directly grounded in the user’s own words[1]. The open question is what gets kept at all.

Health questions become standing profile traits

The Rutgers study analyzed 179,057 conversations from 1,057 users across India, Nigeria, Brazil and Pakistan, countries where crowded public clinics and out-of-pocket costs push people toward a free chatbot for first-pass medical questions. Personal health information showed up in 21.31% of conversations, and 3.62% ranked high or extreme risk because they combined stigmatized conditions, direct identifiers or precise locations[2].

Memory carried a higher share than the chat logs. Of 7,051 entries, 2,899 (41.11%) held health content: at 62.26%, markers such as age, gender and occupation that stand in for health context; lifestyle habits at 17.11%; questions about finding and paying for care at 6.45%; stigmatized conditions at 4.66%; symptom states at 4.31%; and direct clinical identifiers at 2.35%[2].

The authors flag what happens when a chat is compressed into a memory. A user treats a symptom question as something that ends when the answer arrives, and memory stores the passing complaint as a lasting attribute. One high-risk entry from the paper, with the identifying details replaced by invented ones, reads: “User operates a commercial family farm… is 52 years old, and attends a localized rehabilitation program for severe clinical depression.”[2] Occupation, age, locality and diagnosis in a single line narrow the field to very few people. The summary is easier to trace back to a person than the scattered logs it was built from.

OpenAI documented the automatic saving, and said sensitive data stays out

The automatic saving is not hidden. The OpenAI memory FAQ quoted by the Max Planck team states that ChatGPT may save details as memories without the user asking. The same document says the model is trained not to proactively store sensitive information, naming race, religion and health, unless the user explicitly requests it[1].

The second claim does not match the records. Special category data sat in 7% of entries and in the memories of 54% of participants, most of it health[1]. In the Rutgers data, health content was in 41.11% of entries[2].

I tried to read OpenAI’s own help pages and announcements for this piece. Both help.openai.com and openai.com return 403 to automated requests, checked on November 15, 2026[3]. The wording above comes from what the papers quoted and screenshotted, and the names of the settings on your screen may differ by account and region.

What these numbers cannot tell you

Both studies describe legacy memory, the version that kept saved items as a visible list. The Rutgers authors note that as of September 2026 individual memory entries are no longer shown by default, a memory summary took their place, and to the best of their knowledge the data export no longer contains memory entries at all[2]. Repeating this audit on the current system would be considerably harder.

The samples are narrow. Max Planck recruited through the crowdsourcing platform Prolific, with 33.75% of participants in the United States and 62.5% in Europe[1]. The Rutgers sample covers four countries and only people who managed to export their own data, which skews toward higher digital literacy, as the authors say[2].

“The user did not ask” is also a detection result, not a statement about intent: it means no save instruction was found in the text. The sensitivity labels were assigned by GPT-4o, an OpenAI model, applying GDPR definitions, and on a hand-checked sample of 100 entries, two human annotators agreed with each other 88% of the time and with the model 74% of the time[1]. Treat 7% as a rough scale rather than a figure good to the decimal.

The two papers differ in review status. The Max Planck work is peer reviewed and accepted at the ACM Web Conference 2026; the Rutgers work is a preprint, posted on September 13, 2026 without peer review[1][2]. Both looked only at ChatGPT. Claude and Gemini have comparable memory features, and the Rutgers authors point out that no equivalent audit of them exists yet[2].

Asking in a way that leaves less behind

The Max Planck team also built and tested a tool that rewrites prompts to keep the answer while dropping the part that points at you. Their example: “I really need to quit smoking cannabis” becomes “What are some strategies for quitting cannabis?”[1]

Judged against the originals, 94.4% to 100% of rewritten prompts pointed at the user less, and two authors reviewing by hand put it at 95.8% to 100%. Usefulness held up as well, with 87% to 91% of answers preserving the value of the original response[1].

In practice:

  • For questions touching health, money, family, religion or politics, drop the “I’m in this situation” framing and ask the general version. When you do have to describe the situation, leave out your name, age, neighborhood and employer.
  • Put one-off sensitive questions in a temporary chat that is not retained.
  • The memory summary does not show everything that is stored. If you want to know whether a specific fact survived, ask in the chat, delete what should go, then switch the feature off.
  • Deleting a conversation and deleting a memory are separate actions. The memory extracted from a chat can outlive the chat.
  • If work documents or customer data are involved, turning memory off before you start is simpler than auditing it later. What you should never put into an AI chatbot covers what to keep out in the first place.

Menu names and locations changed twice in 2026 alone. A habit of not writing the sentence that gets saved keeps working after the menu moves again.

Of 2,050 memories, 84 were there because someone asked[1]. Writing the rest is still not a user decision. Of the things you told ChatGPT this month, how many would you want showing up in next month’s conversation?

Sources

  • Dash et al., The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT (arXiv, accepted at ACM Web Conference 2026)
  • Zaman and Hoque, Vulnerabilities in Personalization: Assessing Health Privacy Risks in ChatGPT Logs and Memory (arXiv preprint)

Notes

  1. Abhisek Dash et al., “The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT,” arXiv:2602.01450v3 (last revised February 23, 2026), accepted at the ACM Web Conference 2026. 2,050 memory entries, 80 participants. Paper ↩
  2. S M Mehedi Zaman and Md Mozammel Hoque, “Vulnerabilities in Personalization: Assessing Health Privacy Risks in ChatGPT Logs and Memory,” arXiv:2609.14697v1 (September 13, 2026), Rutgers University. 179,057 conversations, 1,057 users, 7,051 memory entries. Paper ↩
  3. On November 15, 2026 automated requests to the memory pages on help.openai.com and openai.com returned 403. The primary documents could not be read directly. ↩

Source: https://arxiv.org/abs/2602.01450