A research team that analyzed the ChatGPT histories of 80 real users found that 96% of the entries in ChatGPT memory were created by the chatbot itself. Only 4% followed an explicit request from the user, according to the Algorithmic Self-Portrait paper on arXiv.
If you have ever opened your saved memories and wondered where a detail came from, this is the first dataset that puts numbers on that feeling. It also shows how much of what ChatGPT keeps about you was never your decision.
TL;DR: Researchers studied 2,050 saved memories from 80 ChatGPT users and found that ChatGPT started 96% of them on its own. About 28% held personal data under GDPR definitions, and 52% held psychological insights about the user. The sample is small, but the memory controls in Settings are worth a look.
What the ChatGPT memory study measured
The authors recruited 80 participants, mostly from the US and Europe, who donated their ChatGPT data using their GDPR access rights. The team then traced each of the 2,050 memory entries back to the message that triggered it, as the paper describes.
That tracing is what makes the 96% figure meaningful. It is not a survey of how people feel, it is a count of who started each memory.
| Measure | Finding |
|---|---|
| Participants | 80 |
| Memory entries analyzed | 2,050 |
| Created by ChatGPT, not on request | 96% |
| Followed an explicit user request | 4% |
| Contained GDPR-defined personal data | 28% |
| Contained special category data | 7% |
| Contained psychological insights | 52% |
Health details showed up despite OpenAI’s stated approach
OpenAI’s own memory announcement says it wants to “steer ChatGPT away from proactively remembering sensitive information, like your health details.” The study found sensitive data in memory anyway.
Special category data, the GDPR term for things like health, appeared in 7% of memories but touched 54% of participants. Health was the most common type, making up 61% of those special category entries, and 35% of participants had health information stored.
Seven percent of entries sounds small until you notice that more than half the people in the sample had at least one. A single stored diagnosis is enough to put a health condition in someone’s profile.
The psychology layer is easy to miss
The researchers also coded memories for what they call Theory of Mind content, meaning notes about what a user wants, intends, feels or believes. That label appeared in 52% of memories and in 97% of participants who had any memories.
Desires dominated at 73% of those entries, followed by intentions at 16%, emotions at 14% and beliefs at 11%. A line like “wants to switch careers” is not a fact you typed in, it is a conclusion the system drew.
Accuracy looked decent. Using GPT-4o as a judge, 77% of memories were directly stated or paraphrased from the user’s messages, while 4% were not supported by anything the user wrote. That 4% is small, but a stored line about you that cannot be traced to anything you said is the kind of error you only catch by going looking.
How much weight 80 people can carry
The authors list their own limits. The sample is small and drawn mainly from US and European users on the Prolific research platform, and it covers only OpenAI’s implementation of memory.
The sensitivity labels also came from GPT-4o rather than from human reviewers alone, though the team checked a portion by hand. Treat the percentages as a strong signal about the pattern, not a precise rate for every ChatGPT account.
The settings worth five minutes of your time
OpenAI’s announcement says memory can be turned off under Settings, then Personalization, then Memory, and that individual entries can be reviewed under Manage Memory. Menu labels change over time, so look for the Personalization section if yours differs.
You can also tell ChatGPT in a conversation to forget something. OpenAI says temporary chats do not appear in history and do not use memory, which makes them the cleanest option for a health question or anything you would not want attached to your profile.
Memory is becoming the central pitch for assistants, as Meta’s Muse Wonder shows, and it is worth reading what yours has written about you. If you want a separate layer of privacy, ChatGPT’s locked chats are one more control to know about, though the study did not examine them.
