Who actually controls what ChatGPT remembers about you?

Analyzing 2,050 memory entries drawn from 80 ChatGPT users who donated their interaction histories under GDPR right-of-access requests, the study reports that "a striking 96% of memories in our dataset are created unilaterally by the conversational system," while only "4% (84) of the memories ... are initiated by participants." The entries came from 1,058 conversations containing 22,971 turns, identified by searching the account's "bio" tool logs for the phrase "model set context updated" and tracing each one back to the user message that triggered it.

OpenAI's own documentation, quoted in the paper, describes two retention modes: an explicit one, where "users have explicitly asked ChatGPT to remember" something, and an implicit one, where "if you share information that might be useful for future conversations, ChatGPT may save those details as a memory without you needing to ask." The authors frame this dichotomy as a gap between the agency OpenAI's policy language implies and what their trace analysis shows happens in practice: the implicit mode appears to account for nearly all memory creation in this sample, so the question of "who holds the agency in shaping and curating this persistent memory" is, empirically, mostly answered by the system rather than the user.

This bears directly on Does ChatGPT shift responses based on inferred political views?, which found the same implicit channel carrying a political persona into unrelated answers; together the two studies describe a single mechanism — unilateral, inference-driven memory writing — with two different downstream effects, content drift and tailored responses. It also sharpens the privacy side of Does chatbot personalization build trust or expose privacy risks?: that longitudinal study found each interaction "ratchets" user expectations, but if most of what gets stored is never requested, expectations may not even register at the point memory is written. It also adds a provenance angle to How does interaction context shape agreement sycophancy in LLMs? — a memory profile shown there to amplify sycophancy most sharply is, per this study, a profile the user mostly did not ask to be built.

The sample is 80 Prolific participants, mostly from the US (33.75%) and Europe (62.5%), pre-screened for heavy ChatGPT use (100+ conversations, 90+ days active) — the authors note this group's "behaviors, and privacy attitudes" may not generalize to the global user base. The study covers only OpenAI's bio-tool implementation as it existed at data-collection time, and the classification pipeline relies in part on "imperfect methodological proxies, including LLMs-as-judges," checked with human oversight but not perfectly validated. What the finding supports is narrower than a universal claim that AI memory is out of users' hands: in this specific, GDPR-donated ChatGPT sample, the large majority of memory-writing events were not user-initiated — enough to question whether OpenAI's stated agency-preserving framing matches observed behavior for this product at this time, not enough to size the gap for other platforms or user populations.