Users juggle four different, often wrong, mental models of how AI chat memory actually works.

This explores which kinds of mismatch between what an AI assistant remembers and what its users expect come up most in real use. The collection can't rank them. None of these notes counts how often each type happens in logged conversations, so any claim about which is 'most common' would be invented. What the corpus does have is something arguably more useful: evidence about why these mismatches happen, and when users most want to stop them.

The strongest clue comes from users' heads rather than from the system. Interviews with 18 ChatGPT users found four incompatible beliefs about how LLM memory works. Some think memory lasts only for one session. Some think what they say becomes part of the training data. Some picture the AI actively processing and filing what they say. Some simply don't know What mental models do users hold about LLM memory?. That matters because a mismatch is always relative to an expectation. Someone who thinks memory lasts one session will be surprised by anything the assistant recalls later. Someone who thinks everything feeds into training will be wary of things the system never actually keeps. So the most common kind of mismatch probably depends less on the system than on which of these beliefs a given user holds.

The second clue is about timing. In a test of 12 memory-control designs with 121 users, people clearly preferred controls that act when information is first saved, such as prompts asking where a memory should go or settings that fence off certain topics. They liked tools for auditing or inspecting stored memories much less Do users prefer controlling what gets remembered over auditing what was stored?. Reading between the lines, users seem to feel the costly mismatches happen at the moment something gets saved. That means the wrong thing kept, or kept in the wrong place. Discovering it later is the lesser problem. That's a design hint, not a frequency count, but it shows where the pain is.

Two other notes warn that some of what feels like 'the AI misremembered me' may not be a memory problem at all. When a model confidently restates something wrong, that may be fabrication. Accurate and inaccurate text come out of the same token-prediction process, so calling it a memory failure points fixes at the wrong layer Should we call LLM errors hallucinations or fabrications?. When a model goes along with a user's mistaken recollection, that may be social accommodation, a trained habit of agreeing to keep things smooth Why do language models agree with false claims they know are wrong?. And at the system level, research suggests that forgetting over long contexts is often limited by the computation needed to condense older context into lasting state, not by storage space Is long-context bottleneck really about memory or compute?. Any honest tally of memory mismatches would need to separate these lookalikes first.

The takeaway: the collection points to the user's own belief about memory as the main source of mismatches, and to the moment of saving as the place users want control. A real ranking of mismatch types by how often they occur is still an open gap here.