Study finds ChatGPT saved memories without explicit requests in 96% of cases

A question asked in passing can become something ChatGPT remembers about a person. Researchers examining donated chat histories found that 96% of analyzed memory entries were created or updated without an explicit request. Gizmodo’s October 22 coverage details how those records included psychological information—and sometimes assumptions rather than facts users had stated.

The team studied histories donated by 80 ChatGPT users who obtained copies through European privacy laws. Researchers identified 2,050 memory entries belonging to 65 participants. The study, presented at the ACM Web Conference 2026, examines what the chatbot retained, not simply what people typed into its conversation window.

From a question to a personal fact

Memory lets ChatGPT carry information from earlier exchanges into later responses. The researchers call the resulting profile an “algorithmic self-portrait”: personalization built from information people disclose in private conversations. Most entries in their sample came directly from those disclosures, but the chatbot also made inferences about the person behind the prompt.

One example shows the distinction. After a user asked for bands similar to Nirvana, ChatGPT recorded that the user liked Nirvana. The request did not explicitly establish that preference. Futurism’s account notes that ChatGPT marked the assumption as weakly supported, preserving uncertainty while still making it part of the stored record.

The records also reached beyond tastes. Researchers classified 52% of entries as containing psychological information, including desires, intentions, emotions and beliefs. Separately, the study classified about 28% as personal data under its interpretation of Europe’s General Data Protection Regulation. These are classifications of stored entries, not percentages of all ChatGPT users.

96%Created or updated without an explicit request

The finding describes memory creation and updates within the study’s sample.

52%Contained psychological information

The category included users’ desires, intentions, emotions and beliefs.

A profile can serve more than conversation

The research predates changes to ChatGPT’s memory system, so its percentages should not be read as a measurement of today’s version. Gizmodo describes a newer approach that synthesizes past conversations into an evolving profile rather than only compiling individual memories. OpenAI says the summary users can inspect may not show everything used to personalize responses.

There is a commercial use for conversational context, too. In an October 21 engineering account, OpenAI says its advertising system can consider nonsensitive information from past conversations when users choose personalized ads. Its ad-selection model interprets the developing exchange and uses tools to retrieve, assess and select potential advertisements.

OpenAI says that process runs separately from ChatGPT’s answer generation and does not influence its answers. That is a boundary around how ads are selected, not a claim that conversational information has no advertising use. Lead author Abhisek Dash told Gizmodo that chatbots’ expanding roles as advisers and companions create privacy risks because people disclose information across many parts of their lives.

Control starts before information is saved

ChatGPT offers several controls, but they address different uses of information. Gizmodo describes options for limiting memory, correcting remembered details and keeping conversations out of future model training:

• Turn off memory in settings to limit what ChatGPT retains through that feature.
• Review and correct the memory summary, while recognizing that it is not necessarily a complete inventory.
• Separately opt out of having conversations used to train future models; that is a different choice from memory.

Deleting a conversation does not necessarily erase information remembered from it, OpenAI says. That distinction makes the retained profile important in its own right: removing the original exchange is not always the same action as removing what the system extracted from it.

The researchers also tested an approach they call an “attribution shield”: asking a general question instead of attaching it to yourself. For example, a person could ask what someone with diabetes should do, rather than identify themselves as diabetic. A small experiment found that such rephrasing could reduce sensitive disclosures, though its effectiveness over longer conversations remains unclear.

Dash argues that responsibility should not rest entirely with users. One change he proposes is asking permission before saving information from a conversation, giving people a decision at the moment a passing disclosure becomes a lasting memory.