Different people picture AI memory totally differently — as a buffer that clears, old training data, or a black box — and that shapes what 'privacy controls' even means to them.

This explores whether different kinds of users want different controls over what an AI remembers about them. The direct answer is that the collection has no study that sorts users into types and compares what each group wants. Read laterally, though, it suggests the more useful split isn't demographic. It's about how each person *thinks memory works*. Interviews with 18 ChatGPT users found four mental models that don't fit together: memory as a buffer that empties after each chat, memory as an extension of training data, memory as active processing of what you say, and memory as a black box. Each model shaped how privacy-aware people were and how they behaved What mental models do users hold about LLM memory?. Someone who thinks memory resets after every chat won't look for a delete button. So 'what controls do people want?' partly depends on 'what do people think is being kept?'

Across those differences, users broadly agree on *when* they want control. A study that had 121 users quickly try 12 memory-control designs found a strong preference for controls at the moment something is saved, like a prompt asking where a memory should go or boundaries set ahead of time. Reviewing logs or inspecting diagrams of why something was remembered ranked lower Do users prefer controlling what gets remembered over auditing what was stored?. Interview work points the same way. People want consent when a memory is created, a way to edit it, limits on how it's used, and a say over anything the system *infers* about them. They actively weigh what they give up against what they get back What memory controls do users actually want from LLMs?. One proposed design splits the work so that users state their preferences and the AI carries out the edits. The idea is to keep the effort of managing memory from wearing users down Does splitting privacy work between users and AI reduce cognitive burden?.

The second split is between types of memory, not types of users. Users expect memories to be organized in three tiers: general knowledge, personal details, and task-specific information. Because no tool shows that structure, heavy users build their own workarounds with separate chat threads, custom prompts and even multiple accounts How should AI agents organize their memory for users?. That suggests control preferences differ by tier as much as by person. You might want one setting for 'my writing style' and a very different one for 'my medical situation.' A related finding shows that sending only 8 of 32 profile fields matched the personalization quality of the full profile, which makes it practical to share a small slice per request and keep sensitive fields on the device Can a smaller user model subset match full model performance?.

The twist is how *personal information gets in* changes how people feel. A 992-person experiment found that personalizing from past conversations made people share more and find the AI less creepy. Personalizing from an upfront questionnaire made people more likely to regret what they had shared Does personalization type shape how people feel about sharing?. Meanwhile, ChatGPT created 96% of 2,050 stored memories on its own, without users asking Who actually controls what ChatGPT remembers about you?. The gap matters. Users say they want control at the moment of capture, yet in practice the system decides what to capture, often without them noticing. On the engineering side, deleting a memory often doesn't truly erase it: retrieval-based systems leaked deleted or unauthorized information in testing Can shared agent memory systems reliably delete information?. That means even the users who most want control may be getting less of it than the interface suggests.