Users have four different stories for how ChatGPT remembers them, and it turns out each one is partly true.

This explores the everyday theories people form about how an LLM remembers (or doesn't remember) what they tell it, and how those theories compare with how memory is really built. The corpus has one direct study on this, and it's revealing. Interviews with 18 ChatGPT users found four mental models that don't fit together. Some see memory as a buffer that empties when the chat ends. Some think what you say becomes part of the training data. Some picture the model actively processing and filing away information about you. And some simply don't know What mental models do users hold about LLM memory?. These models matter in practice. They shape what people are willing to share and how worried they are about privacy.

The surprising part is that none of the four is simply wrong. Each one roughly matches a real layer of how these systems store information. One proposal maps LLM memory onto three brain-like tiers. Model weights act like long-term consolidated knowledge. Retrieval stores act like a fast index of recent experiences. Agentic state (the running notes and goals a system keeps while working) acts like working memory and executive control Can brain memory systems explain how LLMs should store knowledge?. The 'session buffer' user is describing the context window. The 'training data' user is describing the weights. The 'active processing' user is describing retrieval and agentic memory. Each user's model is a partial view of a system with several layers, and it's easy to assume your one layer is the whole thing.

The 'it goes into training data' model also points to something real about how models behave. Models do carry memorized material in their weights, and it affects their answers in ways that are hard to see. One study found that models judge whether a conclusion follows from a premise largely by whether they've seen the conclusion before, not by whether the premise supports it Do LLMs predict entailment based on what they memorized?. So 'what the model memorized' is a strong force even when it isn't 'your conversation becoming training data.'

What no user model seems to predict is that memory can actively hurt the model. One benchmark tested five memory frameworks and found that all five did worse than having no memory at all, sometimes by more than 10%, even when the stored memories were accurate and relevant Can relevant memories actually harm LLM reasoning?. People tend to worry about whether the model remembers, or what it remembers. The less obvious question is whether remembering helps. Engineers are also still inventing new kinds of memory, such as a separate trained 'memory model' that feeds knowledge to a frozen LLM Can a separate memory model inject knowledge without touching the LLM?. Users' mental models are trying to track a moving target.

One honest limit: the corpus has only one study that directly asks users about their mental models, and it's small and comes from a single country. The bigger story here comes from setting that study next to the technical work. User confusion partly reflects real ambiguity, because 'LLM memory' isn't one thing.