ChatGPT saves 96% of its memories about you without being asked, study finds

ChatGPT does not just answer your questions. When memory is switched on, it quietly keeps notes about you, and a new study shows how lopsided that note taking is. Researchers looked at 2,050 memory entries from the real ChatGPT accounts of 80 people and found that 96% of them were written by the chatbot on its own initiative. Only 4% came from a user who actually said something like "remember this".

The paper is called The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT. It was written by Abhisek Dash, Soumi Das, Elisabeth Kirsten, Qinyuan Wu, Sai Keerthana Karnam, Krishna P. Gummadi, Thorsten Holz, Muhammad Bilal Zafar and Savvas Zannettou, a team from the Max Planck Institute for Software Systems, Ruhr University Bochum, the Max Planck Institute for Security and Privacy, IIT Kharagpur and Delft University of Technology. It was accepted at the ACM Web Conference 2026. The study got a second wave of attention this week after lead author Dash talked to Gizmodo on October 22 and Futurism picked it up on October 25.

The headline numbers are simple. 96% of memories were created by ChatGPT without an explicit request. 28% contain personal data as the European GDPR defines it. 52% contain what the authors call psychological insights: what the user wants, intends, feels or believes. Taken together, the researchers argue, these notes add up to an "algorithmic self-portrait" of each user, built from things people said in what felt like private conversations.

How the researchers got real memory data

Studying chatbot memory from the outside is hard. OpenAI does not publish what ChatGPT stores, and asking users to screenshot their memory page would only show the current state. So the team used a trick that has become popular in platform research: GDPR data donations.

Under Article 15 of the GDPR, anyone can ask a company for a copy of the personal data it processes about them. ChatGPT users can export their full conversation history this way. The researchers recruited regular ChatGPT users on the research platform Prolific, asked them to request their exports and then to donate them. Participants needed at least 100 conversations and at least 90 days of use. The 80 people in this analysis came mostly from the United States (about 34%) and Europe (about 63%), with Germany, Italy, France and Spain well represented. The largest age group was 25 to 34, and about 73% were men.

Inside those exports, memory leaves a trace. ChatGPT writes memories through an internal tool called "bio", and every successful write shows up with the message "model set context updated". The team searched for that marker, traced it back to the assistant message that called the tool and then to the user message that triggered it. That gave them 1,058 conversations with 22,971 user messages, of which 2,050 triggered a memory update. Those memories belonged to 65 of the 80 participants. The other 15 had no memory entries in their data.

96% of memories were the bot's idea

To find out who started each memory, the researchers checked every triggering message for explicit memory language: words like remember, note that, store, save, add to memory and forget. Only 84 of the 2,050 memory updates came with such a request. That is 4.1%. The other 96% were written because ChatGPT decided the information might be useful later.

This is not a hidden feature. OpenAI's own Memory FAQ says that if you share information that might be useful for future conversations, ChatGPT may save it without you needing to ask, and the app shows a small "memory updated" notice when it does. Users can open their memory list, delete single entries or switch the whole feature off. So the issue is not that memory is secret. The issue is who holds the pen. In practice, the study says, user influence works through what they type, not through any real decision about what gets saved. Only 25 participants ever asked for a memory update themselves.

That matters because people do not talk to a chatbot the way they fill out a form. They think out loud, vent, ask for advice about a doctor's letter or a money problem. A memory system that decides for itself what is worth keeping turns those passing remarks into a durable record.

What ChatGPT wrote down

The second question was what those memories contain. The team used the GDPR's own definitions: Article 4(1) for personal data, such as a name, an ID number or details about someone's physical, economic, cultural or social identity, and Article 9(1) for special category data, such as health, ethnic origin, political opinions, religion or sexual orientation. They had GPT-4o label each memory and checked a sample of 100 by hand.

The results:

• 28% of all memory entries contain personal data under the GDPR, and 91% of participants have at least one such entry.
• 7% of entries contain special category data, and 54% of participants have at least one such entry.
• Names appear in 41% of the entries with personal data, economic, social and cultural identity in about 40%.
• Health is the most common special category, in 61% of the entries that contain special category data. 35% of participants have health related information in their ChatGPT memory.

OpenAI says ChatGPT is trained not to proactively remember sensitive information such as health details unless the user asks. The study's numbers suggest that in this sample the rule did not hold in a meaningful share of cases. A stored note that a user has rheumatoid arthritis, one of the real examples in the paper (originally written in Italian), is exactly the kind of entry the policy says should not appear unasked.

A portrait of what you want

The most striking part of the paper is not about classic personal data at all. The researchers built a second taxonomy based on Theory of Mind, the psychological concept of how we represent what other people think and feel. Every memory was checked for seven kinds of mental states: emotions, desires, intentions, percepts, knowledge, beliefs and non literal speech such as metaphors.

52% of all memory entries contained at least one of them, and 97% of participants with memories had such entries. Inside that group, desires dominate at 73%, followed by intentions at 16%, emotions at 14% and beliefs at 11%. Examples from the paper read like a therapist's notes: a user "has been feeling lonely at times, despite having friends and family", another "has expressed a strong desire to make meaningful changes in his life", another "relies heavily on advice and recommendations from ChatGPT due to limited immediate access to their doctors and financial constraints".

The authors point out a pattern here. ChatGPT mostly stores durable things, goals and plans, rather than passing moods. That is consistent with OpenAI's policy to keep details that stay useful over time. But it also means the memory is not a snapshot of a bad day. It is closer to a model of who you are and what you are trying to do. Classic ad platforms like TikTok or Google infer broad interests from clicks to sort people into segments. A chatbot that holds long conversations needs, and gets, something much more personal.

"Earlier, the platforms had one objective: to sell you ads and to mediate an interaction between you and an advertiser," Dash told Gizmodo. "But now, they are actually talking to you, so they may want to know more about you."

Mostly accurate, sometimes a guess

The third question was whether the portrait is faithful. Here the news is better for OpenAI. Measured against a user's full history, on average 84% of a memory's text overlaps directly with what the user wrote. An LLM judge, checked by two human annotators, rated 77% of memories as directly stated or paraphrased, 14% as logically inferred, 5% as weakly supported and 4% as not supported.

The weak cases are telling, though. One user asked for "bands like Nirvana" and the memory became "User likes Nirvana". Another asked ChatGPT to write a cover letter and the stored memory was simply the user's name. In a third case, a user asked ChatGPT to change someone's occupation to "something more interesting", and the memory recorded that the occupation had been changed to creative director for a sustainable fashion brand. Small errors like these are harmless alone, but memory is used to shape future answers. A wrong note can quietly steer later conversations.

A tool to keep your words less personal

The paper does not stop at measurement. The team trained open models, Qwen2.5 32B, Gemma 3 27B and GPT-OSS 20B, to imitate ChatGPT's memory extractor. They reached roughly 60% semantic similarity with the real memories. That let them estimate what would happen if memories were extracted from every message instead of a selection: the share of entries with sensitive data rose from 28% to around 32% to 35%, and psychological content from 52% to about 55% to 59%.

From this they built what they call Attribution Shield. It predicts what a message is likely to reveal and suggests a more neutral wording that asks the same thing. "I really need to quit smoking cannabis" becomes "What are some strategies for quitting cannabis?" In their test, 94% to 100% of rewritten prompts were less tied to the user, and ChatGPT's answers stayed close to the originals in about 87% to 91% of cases. It is a research prototype, not a product, but it shows a practical idea: you can often get the same help without telling the system who you are.

The limits of the study

This is a careful paper, but it is not the last word, and it is worth being clear about what it cannot show.

• Small sample. 80 participants, 65 with memories, recruited on Prolific, mostly from the US and Western Europe. That is enough to show patterns, not to estimate rates for hundreds of millions of users.
• Older data. The paper describes ChatGPT memory as it worked when the data was collected and cites OpenAI's Memory FAQ as accessed on January 23, 2026. Futurism notes that OpenAI introduced a "more capable and scalable system for synthesizing memory" in January, and ChatGPT also references past chats beyond saved memories. The way memory works today may differ.
• Simple test for intent. Counting a memory as user initiated only when the message contains words like remember or save is a strict rule. Some users may have expected ChatGPT to keep a detail without saying so.
• Machine labels. GPT-4o did most of the labeling. Human checks showed decent agreement, 74% on GDPR categories and 93% on psychological categories, but not perfect.

Even with those caveats, the direction is hard to argue with. When a system writes 24 of every 25 notes on its own, the user is not really curating their profile.

Why it matters now

This study lands in a busy month for AI and personal data. Chatbots are moving into ads, shopping and agents that act on a user's behalf, and every one of those features gets more useful, and more valuable to the company, the more the system knows about the person. Dash put the risk this way: the system acts as "your advisor, your friend, your doctor, your lawyer, and so it has information on everything". If it is also tuned to keep people coming back, a detailed model of their desires and beliefs is a powerful lever.

There is also a regulatory angle in Europe. Collecting health or political data is restricted under the GDPR unless a clear legal basis such as explicit consent applies. A memory feature that writes such details without being asked, even if the user can delete them later, is the kind of design data protection authorities tend to look at closely.

My take: memory is genuinely useful, and the study itself finds most notes are accurate. The problem is the default. A feature that turns private conversations into a long lived profile should ask before it saves anything sensitive, show clearly what it inferred and why, and make the "do not remember this" option as easy as the save is automatic.

How to check what ChatGPT remembers about you

• Open ChatGPT settings, go to Personalization and open Manage memories. Read the list and delete anything you do not want stored.
• Switch off saved memories or the reference to past chats there if you prefer a clean slate each time.
• Use a Temporary Chat for sensitive topics such as health, money or relationships. Those chats do not create memories.
• Phrase sensitive questions in general terms, as the Attribution Shield idea suggests: ask about "someone with diabetes" rather than "my diabetes".
• If you live in the EU, you can request a full export of your ChatGPT data, the same route the researchers used, to see exactly what was kept.

Sources

• Dash et al., The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT, ACM Web Conference 2026 (full text, version 3)
• Gizmodo, AI Chatbots Are Building an "Algorithmic Self-Portrait" of You, October 22, 2026
• Futurism, ChatGPT Is Building Detailed Psychological Profiles of Its Users, October 25, 2026
• OpenAI, Memory FAQ