People often use the word memory for several different mechanisms in AI systems. The analogy is convenient, but it can obscure more than it explains.
A model’s learned parameters, the current context window, saved user information, and externally retrieved documents are not the same thing.
One word, several mechanisms
A useful mental model is to ask two questions: where is the information stored? and how does it get into the model’s current context? Those questions make most apparent mysteries much easier to reason about.