LLMs suffer from the problem of the Context Window, and that any information has to be provided as background within the context, and it doesn’t remember anything from previous generations.
In some way, building a system with the LLM in which a community of humans guide the output of the LLM over time is a form of long-term memory that the context window cannot provide. The humans queue material generation, and do the long-term planning for the LLM.