What mental models do users hold about LLM memory?

A thematic analysis of semi-structured interviews with 18 Chinese participants, centered on ChatGPT's memory feature as a case study, found that users hold four distinct and often inaccurate mental models of how retrieval-augmented-generation (RAG) based memory works. Mental Model 1 treats memory as "a transient, dialogue-specific buffer" that resets once a conversation window closes — P1 stated flatly that "the LLM's memory can only be maintained in the same dialogue window," and P5 agreed memory "exists in the same dialogue, but not across dialogues." Mental Model 2 treats memory as an extension of the model's core training data: P3 wondered if memory worked "through training on massive data," describing it as a "long-term memory model ... built upon the original training data ... comprehensively collecting new information ... then conducting further training," and P16 asked whether it meant "taking past chat data, re-labeling it, and then put[ting] it into its model for training." The excerpt names a third model, memory as an active information-processing mechanism, and a fourth: a subset of participants who simply acknowledged they did not understand how it worked at all.

The study's authors frame these mental models as consequential rather than merely a curiosity of comprehension, because the model a person holds "largely influence[s] whether [they] could be aware of the potential privacy risks and could effectively control" memory. A user running Mental Model 1 has no reason to worry about a persistent profile accumulating across sessions; a user running Mental Model 2 worries instead that private conversations are being folded into a "public knowledge base" accessible to other users or future training. The models also drive behavior directly — participants holding Model 1 described opening new chats specifically to "reset" the system when it got stuck on a flawed instruction, a workaround that only makes sense under that model's premise.

This sits alongside Does chatbot personalization build trust or expose privacy risks?, which shows privacy concern rising alongside trust as personalization deepens over time; this interview study supplies a mechanism for why that concern is so unevenly distributed across users — it depends on which of four incompatible stories a person tells themselves about what memory even is. It also qualifies Do language models know what they don't know about users?: that note addresses the system's blind spot about the user, while this one shows the user's own blind spot about the system is just as structurally consequential, and in the excerpt is explicitly named by participants as a "lack of epistemological certainty."

The excerpt does not establish how these four models are distributed in a general population — the sample is 18 participants from one country, recruited online, discussing one product (ChatGPT) as an anchor case, so prevalence claims beyond this group aren't supported. It also stops short of the paper's full RQ2/RQ3 findings on protective strategies and design demands, which the companion note What memory controls do users actually want from LLMs? covers. The implication the evidence does support: any interface explaining "what the AI remembers" is working against at least four different starting assumptions in its audience, not one.