To the extent it "knows" (using that word loosely) about the persona, it's deriving that information from its system prompt. The model itself has no awareness.
The sooner we stop anthropomorphizing AI models, the better. It's like talking about how a database is sentient because it has extremely precise memory and recall. I understand the appeal, but LLMs are incredibly interesting and useful tech and I think that treating them as sentient beings interferes with our ability to recognize their limits and thereby fully harness their capability.