By learning it just creates a list of “facts” it extracts from the conversation.

There’s no updating model weights.

(Though I suspect you know that due to your use of quotes this is not for posterity)