That's encoded in the model weights, not "memory". Basically, there is no context outside of the context that you give the model. When you ask it a question, th

That's encoded in the model weights, not "memory". Basically, there is no context outside of the context that you give the model. When you ask it a question, those model weights don't change. It doesn't "remember" what you asked. This is why chain-of-thought reasoning works so effectively: it lets the model "use" the context as a sort of scratch pad to build up a response. Without it, the model isn't capable of mimicking thought because it's only capable of predicting based on the current context.