Does knowing how AI works change how you feel about it remembering things about you?

This explores whether technical know-how changes how people make sense of an AI that remembers or infers things about them. The short answer is that the collection has no head-to-head study of novices versus experts. What it does have is more interesting: evidence that people don't share one picture of what AI memory even is. In interviews with 18 ChatGPT users, researchers found four mental models that don't fit together. Some saw memory as a buffer that only lasts for one session. Some saw it as an extension of the training data. Some saw it as active processing of what they say. Some simply didn't know What mental models do users hold about LLM memory?. Those models shaped how careful people were about privacy. So the real split may be less about 'technical vs. non-technical' and more about which picture of memory someone happens to hold.

Here's the catch: the expert's picture can be wrong too. A technically fluent user might reason from what they know about models: weights are frozen, the context window is limited, so nothing persists. But How do AI products decide what to remember between conversations? points out that memory is a product design choice, not a property of the model. The app decides which stored memories get pulled into each request. Knowing how transformers work doesn't tell you what a given product keeps or brings back. Someone who reasons from model architecture can be as miscalibrated as someone who reasons from intuition. Their errors just point in a different direction.

What users want from memory may be more consistent than what they understand about it. In a study where 121 people quickly compared 12 memory-control designs, people strongly preferred deciding up front what gets remembered over auditing records afterward Do users prefer controlling what gets remembered over auditing what was stored?. That points to a design lesson that doesn't depend on expertise. If people hold such different mental models, controls at the moment something is saved work no matter which model the user carries. Inspection tools, like causal graphs of stored memories, quietly assume a level of understanding many users won't have.

There's one more twist that neither group is likely to expect: remembering accurately isn't the same as remembering helpfully. Memories that were recorded correctly and were relevant to the task still dragged reasoning below a no-memory baseline in every framework tested Can relevant memories actually harm LLM reasoning?. What a memory is used for, such as clarifying versus just being related, matters more than whether it's on topic Does retrieved memory quality depend on its functional role?. Both novices and experts tend to judge memory as a privacy question ('what does it know about me?'). The research suggests an equally important question is 'how is what it knows bending its answers?'

If you want a direct comparison by expertise level, the corpus doesn't have one yet. That's a real gap, not a settled question.