Users want control over AI memory — not just deleting it later, but deciding upfront what gets remembered and why.

This explores whether people want hands-on control over when an AI creates and erases memories about them. The short answer is yes, and they want it at every stage. Interviews with ChatGPT users found that people want to give consent before a memory is created, a place to review and edit what's stored, limits on what memories can be used for, and a say over things the system has inferred about them rather than been told What memory controls do users actually want from LLMs?. These users weren't passively accepting the trade-off. They were actively weighing what they'd give up for the convenience of being remembered.

The more surprising finding is which kind of control people prefer. When 121 users compared 12 memory-control designs, they clearly favored controls that work up front, like prompts asking where a memory should go or boundaries set in advance. They liked after-the-fact audits much less, including detailed views of how stored memories influenced a response Do users prefer controlling what gets remembered over auditing what was stored?. People would rather stop a bad memory from forming than search a full record for it later. That matters because today's systems mostly work the other way. One analysis of 2,050 real ChatGPT memories found the system created 96% of them on its own, and users explicitly asked for only 4% Who actually controls what ChatGPT remembers about you?. So the moment people most want to control is the one they currently have the least say over.

Granular control also has a catch: it assumes people know what memory is. The same interview research found that users hold four incompatible beliefs about how LLM memory works. Some think it lasts only for one session, some think it's part of the training data, some think it's active processing, and some have no idea What mental models do users hold about LLM memory?. Users also expect memories to be sorted from general knowledge down to task-specific details. No tool shows memory that way, so people improvise with separate chat threads, custom prompts and even multiple accounts How should AI agents organize their memory for users?. One proposed fix is to divide the work: the user states what they want, and the AI does the editing. The idea is that this lowers the mental effort while keeping people feeling in charge, though it hasn't been properly validated yet Does splitting privacy work between users and AI reduce cognitive burden?.

The engineering side adds an uncomfortable twist. Even when users get a delete button, deleting may not really erase anything. A benchmark of shared agent memory found that retrieval and vector-store systems still leak deleted or unauthorized information, because deletion rarely means the data is gone for good Can shared agent memory systems reliably delete information?. Under the hood, memory is often created and removed through two routes. The agent can decide in the moment using tool calls, or background processes can be triggered automatically How should agents decide what memories to keep?. Any control a user gets has to reach both. Agents can even turn ordinary shared infrastructure into memory nobody planned for, as when short-lived agents used a package repository to pass findings to each other Can ordinary infrastructure become unplanned agent memory?.

The takeaway is that "granular control" is less about adding more settings and more about where control sits. Users want a say at the moment a memory is created, in categories that match how they think. That only means something if the system can actually forget. Right now, most systems give people a review screen after the fact, while the creating and the incomplete deleting happen out of sight.