Arguably it can be done now, you just need to take the architecture of gpt3 plus davinci and allow it to process enough tokens to write a full book without forgetting things from the start. I actually worked with it to create a concept for a game idea I had but after a certain amount of words, it gets confused. I was able to adjust for this by giving it summaries of where the story was upto and reminding it of points that we had already incorporated.

So tldr, 1) allow it to process more tokens, or 2) compensate for this with human guidance.