You're vastly overestimating LLM capabilities and underestimating what it takes to get them to both stick with a functionality like this and ensure they crap out something that will keep a player engaged for more than a few minutes. To get one to play-act in a way that you could use in a game you'd have to aggressively train it yourself - and that means having your own. If you just embed a call to ChatGPT or something it'll very obviously sound like ChatGPT and won't 'remember' any of your game parameters or what the player has said to it in the past. The results will be mediocre-to-offputting at best and the amount of time you'd have to spend fussing with it to achieve any level of consistent quality won't be worth it.
If you're just imagining making a quick buck with some scammy shovelware you might get some takers, but if you're hoping that you can leave most of the heavy lifting to the AI and end up with a good result you're going to be disappointed.
You are understimate modern AI langue models. The size of context has increased over the year and now it can hold text more than 20 pages (128k tokens). I roleplay with different LLM and I can talk to them at least 30 minutes on my local machine. But generaly you are right. That's why I'm interested in a building game like Feelings Training because it has text fragments divided by dates. And there are a few tricks that will help retain all the contextual information throughout the game.
And I used chatGPT as an example of LLM, I prefer another one uncesored from META. As for remembering players responses, I can save the context of the responses, not sure how exactly it will work throughout game, experiments are needed.
Generally what you are talking about is mostly technical details. For example imagine you take care of a not very clever girl. You can't expect from her to remeber all details of your life. She probably good at remebering the current context of dialog but can forget some details of the previous dialog. It seem only natural.
And about "quick buck" from one point you are right, I am also looking for the best investment/return ratio. But I pro and experienced in developing this types of games and can provide a high quality. But I understand many peoples very confused by products with AI. They see it as press button "make game" and that's it. It's actually the main problem.