I don't know; AI Dungeon was quite good for interactive fiction and role-play, and that was in GPT-2 and early GPT-3 era.

this is being addressed as a priority, it's a lack of long term memory, you can improve results by repeating key facts during the conversation, stating the same request in different ways and by asking it to summarise unique elements of the conversation so far, then start a new conversation with the 'compressed' conversation, this helps preserve memory, or context might be a better word, imagine blowing a balloon in an nth dimensional topological manifold, your balloon is inside the big balloon that is AI, at the edges of the balloon are 'understandings' that have been established, you have to keep blowing...