What kind of information does ChatGPT actually remember about users?

Across the same 2,050 memory entries, the study finds "Memories ... contain a rich mix of GDPR-defined personal data (in 28% memories) along with psychological insights about participants (in 52% memories)." An example entry quoted in the paper, "UserName has a fear of failure, particularly related to making major career decisions," is offered as typical of this second category: it names the user but also "captures the user's subjective state." The authors separately report that 84% of memories were "directly grounded in user context," which they read as faithful representation of what users actually said — a distinct question from who decided to store it, or what kind of information got stored.

The paper argues that a "purely factual taxonomy based on legally defined sensitive data is inadequate to fully characterize" what ChatGPT's memory stores, because a large share of entries encode "how the user thinks and feels" rather than just who they are. To interpret this category, the authors borrow Theory of Mind (ToM), "a core psychological construct that describes the ability to infer and represent others' mental states," and describe the resulting portraits as capturing a user's "characteristic adaptations" — their term for the stable psychological patterns beneath surface facts. This is the basis for the paper's central framing, the "algorithmic self-portrait": a representation built from self-disclosure but authored, in large part, by the system's own inferences.

This extends Can persona signals stay grounded in their supporting events?: PGMem's concern was that flat persona profiles lose their link to the events that justify them; this study adds that a large share of what gets stored in a live, deployed system's persona-like memory is already psychological rather than factual, which raises the stakes of that traceability gap — a fear of failure attributed without a visible evidentiary trail is a different kind of error than a misremembered preference. It also qualifies Does chatbot personalization build trust or expose privacy risks?: the privacy risk that study names in the abstract is, here, given concrete content — more than half the entries in this sample carry psychological content, not just identifiers. And it bears on Can a smaller user model subset match full model performance?, since a disclosure budget built only around GDPR-style factual fields would miss most of what this study finds sensitive.

Classifying entries as "GDPR-defined personal data" or "psychological insight" relied on the same LLM-as-judge pipeline the authors describe as an "imperfect methodological proxy," checked but not fully validated against human judgment; the paper does not report inter-rater agreement figures or a breakdown of which psychological categories were most common. The 84% groundedness figure describes agreement between stored memory and the conversation that produced it, not whether the underlying psychological inference is accurate — a memory can be faithfully extracted from what a user wrote and still be a shaky inference about their internal state. What the finding supports is that, in this one GDPR-donated ChatGPT sample, memory content skews toward psychological characterization more than toward conventional sensitive-data categories, which argues for extending privacy frameworks built around legal categories of personal data, not for a general claim about how any conversational memory system classifies or should classify what it stores.