What your AI remembers about you, without asking you
Stéphane Commenge, with Claude ·
The Fact
Researchers analyzed ChatGPT's memory using real data from 65 users, or 2,050 memories. 96% were created on the assistant's own initiative, 28% contain personal data and 7% contain sensitive data as defined by the GDPR.
The Signal
Assistant memory saves time on projects that last.
But it fills up behind the scenes: the AI decides on its own what to keep, including sensitive data, and few users open that memory to look at it.
The Action
This week:
1) Open the memory of the assistant you use most, in its settings.
2) Read what it contains and delete whatever is wrong, outdated or too personal.
3) Check whether you can export it, so you can take it with you the day you change tools.
Series "AI memory", episode 1
"Remember that."
How many times have you said it to your AI? With ChatGPT, out of 25 memories it keeps about you, only one comes from that phrase. The other 24, it decided on its own to keep.
Assistant memory is real progress. It picks a subject back up without everything being explained again, knows your preferences, follows a project over time. One question is rarely asked: who decides what gets kept, and who looks at it?
A study that opens the box
Researchers from two Max Planck institutes (for Software Systems and for Security and Privacy), Ruhr University Bochum, IIT Kharagpur and TU Delft looked inside ChatGPT's memory. On real data.
The method is simple and solid. Participants exercised their right of access under the GDPR, retrieved their data, then handed it over to the researchers. Of 80 participants, 65 had saved memories: 2,050 memories, tied to 1,058 conversations.
The study, "The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT", was accepted at the ACM Web 2026 conference. The press picked it up on October 26.
What they found
The AI decides on its own, or nearly. 96% of memories are created on the assistant's own initiative. 4% answer an explicit request ("remember that...").
Personal data. 28% of memories contain some, as defined by the GDPR. 91% of participants have at least one in memory.
Sensitive data. 7% of memories fall under the GDPR's special categories, health first. 54% of participants have at least one.
A psychological portrait. 52% of memories describe a mental state: what the user wants, believes, feels. Desires dominate.
Faithful memories. 84% correspond directly to what the user wrote.
The danger is not that the AI makes things up. It's what it chooses to keep, without you seeing it.
The researchers speak of an "algorithmic self-portrait". A portrait nobody commissioned, painted over the course of conversations. Yet many users confide in their assistant as they would in an advisor, a friend, a doctor or a lawyer, as the press article that relayed the study points out.
A memory nobody looks at
Memory is not the problem. On a project lasting several months, it saves time, provided the AI goes and finds the right information at the right moment. The problem is a memory that fills up behind the scenes and that we rarely open.
Three questions.
- Do you know what it contains? With many assistants, memory can be viewed in the settings. Do the exercise once. I did it with the AI I use every day. In its notes on my projects, lines that had become false slept next to the ones that had replaced them. Nothing said which one still held. I've seen better-kept paper files.
- Can you correct it? A false or outdated piece of information keeps influencing the answers.
- Where does it live, and can you take it back? A memory locked inside a single platform doesn't follow you when you change tools.
Your AI knows a lot about you. And you, what do you know about it?
What I'm building with Mycelora
This is the question Mycelora started from, the memory I'm building at EvidencAI.
Yours. Each memory can be read, corrected, pinned, archived in your dashboard. The whole thing can be exported.
Filed. A type (decision, pitfall, abandoned idea, landmark, state), a date, a project and, if needed, an expiry condition.
Hosted in Paris at Scaleway. Database and functions. The AI models that organize it also run at Scaleway.
Shared. The same memory for Claude, ChatGPT, Perplexity, Mistral Vibe and any AI that accepts a remote MCP connector.
The beta is open: mycelora.ai
Next episode
This past summer, the CNIL (France's data protection authority) and the Conseil de l'IA et du numérique (France's AI and Digital Council) published a note on the memory of AI agents, listing six GDPR principles put to the test. We'll come back to it.
Sources
- Dash, Das, Kirsten, Wu, Karnam, Gummadi, Holz, Zafar, Zannettou, "The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT", The ACM Web Conference 2026: arxiv.org/abs/2602.01450
- Unilad Tech, October 26, 2026: What ChatGPT knows about you as researchers uncover detailed psychological insights