Could this be used to train an LLM? It seems the hidden states could be used to learn how to store history.