The token limit consists of about 1500 words. It's not entirely clear what happens when it reaches that limit. In GPT3, it just stops being responsive after 1500 words. For ChatGPT, I think it may just progressively dump its memory to avoid the session being interrupted, although, I've also had it getting network errors or just getting really slow at responding after a long conversation, so I'm not entirely sure. The point is that it stops working as intended after around 1500 words or at the very least forgets things said 1500 words prior.