ChatGPT shows you a tidy list of what it remembers — but can't tell you how it actually decided what to keep.

This explores why a memory panel that lists stored facts falls short of real transparency. The corpus points to one main reason: a list shows the *contents* of memory, while users are confused about the *process*. When researchers interviewed 18 ChatGPT users, they found four mental models of what memory even is, and the models don't fit together. Some users think memory only lasts for the session. Some think it becomes part of the training data. Some think the system is actively processing what they say, and some have no idea What mental models do users hold about LLM memory?. A list of saved items can't settle which of these is right. Someone who believes their chats feed into training will look at a short, harmless list and still not know whether the rest of what they said went somewhere else. The panel answers 'what's saved?' when the reader's actual question is 'what is this thing doing with me?'

The second gap is timing. Viewing a list is an after-the-fact audit, and users don't seem to want that much. In a speed-dating study, 121 people compared 12 memory-control designs. They clearly preferred controls at the moment something gets captured, such as prompts asking where a memory should go or boundaries on what can be remembered. Passive inspection ranked lower, even when it was elaborate, like graphs showing why a memory influenced an answer Do users prefer controlling what gets remembered over auditing what was stored?. The interview work supports this: users want consent when a memory is created, limits on what it can be used for, and a say over *inferred* information, not just the things they said outright What memory controls do users actually want from LLMs?. A panel covers one stage of that lifecycle. It skips creation, it doesn't cover what memories are used for, and inferences often never show up as readable entries.

The part you may not have expected to care about: even a perfectly accurate panel can mislead you about how memory affects answers. MemTrapBench tested five memory systems. Memories that were stored correctly and were relevant to the task still made reasoning worse than having no memory at all, with drops of more than 10% Can relevant memories actually harm LLM reasoning?. Related work finds that the *role* a memory plays in a conversation, for example clarifying a constraint versus adding noise, matters more than whether it's on-topic Does retrieved memory quality depend on its functional role?. So you could check every entry, find them all true, and still be getting answers bent by them. The problem shows up when memory is used, and a list can't show that.

What might work better? Two design ideas in the corpus put memory *in the workspace* instead of in a settings page. JarvisHub puts prompts, references, versions, and feedback on a shared canvas that both the user and the agent can see, so memory is part of the visible work rather than hidden state Can a shared canvas serve both human and agent memory?. Lam and colleagues' 'just-in-time objectives' approach shows the system's choices at each step of the interaction Does visible design structure keep HCI expertise necessary as AI advances?. Both suggest that transparency should be visible in the moment, not looked up later. One caveat: no study in the corpus tests memory panels directly. This answer is pieced together from studies of user mental models, user preferences, and how memory is used.