Which stored memories actually shape each response?

The same CHI 2025 Extended Abstracts study by Jones, Stemmler, Su, Kim and Kuzminykh — six interviewed users of ChatGPT, Claude, Copilot and similar tools, plus 54 Reddit threads — finds a desire distinct from simply wanting to know what is stored: participants wanted to know how stored memories causally shape a given answer. P3 put it directly, wanting "to know which parts of the memory it has about me go into, like, which responses, and to understand more about how it decides which portions [...] of that list [of memories] is like... you know, [it is] 100 or 200 items... like, that's a lot of things to remember about me. And to decide, like, which parts of that list actually go into a response versus not." P6 described the same gap from the opposite side: "I can never say for sure whether it remembers this thing or not. And even sometimes I tell it to remember, it may forget."

The authors frame this as a layer of opacity beyond simple content-awareness: users weren't only unsure what the system stored, they were unsure how storage converts into behavior. The excerpt's sharpest illustration is P3 asking ChatGPT to guess their Myers–Briggs personality type: "it got three out of the four [indicators] correct. So like that, that kind of threw me off, because I don’t talk much about my personality" and "ChatGPT guessed [my personality], just based on regular interactions that were mostly about work and research." Other participants described a similar jolt on first opening their memory panel — "I didn't notice it, but now I'm looking at the memory, it's definitely collected a bunch of information about me... It's scary" (P1); "Yeah, it remembers quite a lot. Oh, my God" (P2) — and the paper reports that "most participants did not want the system to remember such personal details over a longer term" once they saw how far inference reached.

This extends Do users notice when LLMs infer and retain private information? one step further: that study's participants didn't know specific private items were stored at all until told; this study's participants already knew the system stores things but still lacked any way to see which stored items drove which output, so disclosing content isn't the same as disclosing influence. It also adds a concrete trait-level case — the MBTI guess — to the pattern in Does ChatGPT's memory feature feel like a relationship or a machine?: an inference the user never stated directly reads as a sharper trigger for that negative reaction than a memory the user recalls sharing.

The excerpt doesn't establish how often inference like this occurs or how accurate it generally is — the MBTI guess is one participant's single anecdote, not a measured rate — and the sample is six people who rated their own LLM expertise at 4/5 while two rated their technical understanding of memory mechanics as low as 2/5 and 1/5, so the gap between perceived and actual understanding of inference likely runs deeper in a less expert population. The implication the evidence supports at this strength: a memory panel that just lists stored items is answering the wrong transparency question for users like these — the one they're actually asking is which items get used, and how far the system infers beyond what was ever typed.