Researchers found that 96% of the ChatGPT memory entries they examined were created by the chatbot, not explicitly requested by users. Their study also found personal data in 28% of entries and psychological insights in 52%, putting a concrete measure on what conversational memory can retain.
OpenAI Watch analysis
What happened
The study examined 2,050 memory entries associated with 65 ChatGPT users. Futurism reports that the researchers analysed donated chat histories from 80 people who obtained their records through European data-protection laws. The difference between 80 donors and 65 users with identified memory entries matters: the findings are about the material the researchers could examine, not every ChatGPT user.
The researchers found that some stored memories reflected assumptions rather than direct statements. In one example, ChatGPT recorded that a user liked Nirvana after they asked for bands like Nirvana, even though the user had not said so outright. The article also notes an important limitation: OpenAI has since updated its memory system, so these results do not necessarily describe the current version. Read Futurism’s report.
Why it matters
Memory can save people from repeatedly explaining their preferences or circumstances. But these findings suggest the system may also build a profile from conversational clues, including details users did not expressly ask it to remember. That shifts the privacy question from “What did I tell the chatbot?” to “What did it infer, and what did it keep?”
The study does not establish how often these findings occur across all users, nor whether the stored information led to a particular harm. It does offer a useful snapshot of what researchers found in a defined set of records, and why user control over memory deserves more than a glance at the settings menu.
Our read
Personalisation is useful; a system quietly drawing conclusions about you is a different proposition. The striking finding is not simply that ChatGPT remembers details, but that the researchers say most entries were created unilaterally. If you use memory, check what is stored and whether the controls match what you expect. A chatbot remembering your writing preferences is one thing. Making a confident note from a hint is another.
What to watch
- Whether follow-up research examines newer versions of ChatGPT’s memory system.
- Whether OpenAI makes stored memories and their origins easier for users to inspect and manage.
- How often chatbot memories contain inferred, sensitive or inaccurate information.
Discussion spark: Should chatbot memory be limited to details users explicitly ask it to save, or is useful personalisation worth allowing systems to draw and retain their own inferences?
Sources and evidence
- Source update (25 October 2026, 11:55 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.