ChatGPT Memory Study: 96% of Saved Memories Not Requested
Researchers looked at the ChatGPT conversations of 80 real users. They found that ChatGPT created 96% of the saved memories by itself. Only 4% were saved because users clearly asked ChatGPT to remember something. These findings come from a research paper called Algorithmic Self-Portrait, published on arXiv.
If you have ever looked at your ChatGPT memories and wondered why something was saved, this study helps explain it. It is the first dataset to show with numbers how often ChatGPT saves information about people without being asked.
In short: Researchers looked at 2,050 saved memories from 80 ChatGPT users. ChatGPT decided to save 96% of these memories on its own. About 28% contained personal information under the European privacy law called GDPR. Another 52% contained information about users' thoughts, feelings, or wishes. The study included only a small number of people, but its findings give users a good reason to check their memory settings.
What the researchers looked at
The researchers found 80 people willing to share their ChatGPT data. Most lived in the United States or Europe. They used their rights under the GDPR privacy law to get copies of their data. The researchers then looked at 2,050 saved memories and found the original messages that caused ChatGPT to save each one. The research paper explains how they did this.
This is why the 96% number matters. The researchers did not simply ask people how they felt about ChatGPT memory. They counted how many memories ChatGPT created on its own and how many users specifically asked it to save.
| 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% |
ChatGPT saved health information even though OpenAI said it tries to avoid doing that
In its announcement about memory, OpenAI said it wanted ChatGPT to avoid saving sensitive information, such as health details, unless users asked it to. But the researchers found that ChatGPT had saved sensitive information anyway.
Under the GDPR privacy law, certain sensitive information, including health details, is called special category data. This kind of information appeared in 7% of saved memories. But 54% of the people in the study had at least one memory containing this type of information. Health information was the most common kind. It made up 61% of these sensitive memories, and 35% of participants had health information saved.
Seven percent might not sound like much. But more than half of the people studied had at least one memory containing sensitive information. Even saving just one medical diagnosis means ChatGPT has added a person's health condition to what it remembers about them.
ChatGPT also remembers what it thinks about your mind
The researchers also looked for memories about what people want, plan to do, feel, or believe. They called this Theory of Mind information. They found it in 52% of saved memories. Among participants who had saved memories, 97% had at least one memory containing this kind of information.
Of these memories, 73% included things people wanted, 16% included their plans or intentions, 14% included their emotions, and 11% included their beliefs. For example, ChatGPT might save a note saying someone wants to change careers. That note may not be something the person directly told ChatGPT to remember. Instead, ChatGPT may have worked it out from the conversation.
Most memories seemed fairly accurate. The researchers used GPT-4o to check them. It found that 77% of memories were based directly on something users said or were written in different words with the same meaning. But 4% were not supported by anything the users had written. Four percent may seem small, but a wrong memory about you could stay saved unless you check your memories and notice the mistake.
What a study of only 80 people can tell us
The researchers explained that their study has limits. They looked at only 80 people, mostly from the United States and Europe. These people were recruited through a research website called Prolific. The study also looked only at how OpenAI's memory system works, not memory systems made by other companies.
The researchers also used GPT-4o to decide which memories contained sensitive information. Human researchers checked some of these results themselves, but they did not check every memory by hand. So the numbers give us useful evidence about how ChatGPT memory works, but they do not tell us the exact percentages for every ChatGPT user.
Memory settings you can check in five minutes
OpenAI's announcement says users can turn off memory by opening Settings, choosing Personalization, and then selecting Memory. It also says users can review individual saved memories under Manage Memory. The names of these settings can change, so look in the Personalization section if your menu looks different.
You can also ask ChatGPT during a conversation to forget something. OpenAI says temporary chats do not appear in chat history and do not use memory. This makes temporary chats a useful choice when asking about health or discussing something personal that you do not want saved as part of your profile.
Remembering information about users is becoming an important feature of AI assistants. Meta's Muse Wonder is another example. That is why it is worth checking what ChatGPT has remembered about you. If you want another way to protect your privacy, you can also learn about ChatGPT's locked chats. However, the researchers did not study that feature.