It lacks cohesion the longer it goes on, not so much "hallucinating" as it is losing the thread, losing the plot. Internal consistency goes out the window, previously-made declarations are ignored, and established canon gets trounced upon.
But that's cuz it's not AI, it's just LLM all the way down.
Depends on complexity and the number of elements to keep track of, and varies between models and people. Try it out for yourself to see! :)