I've discovered that if you have a long enough chat with #ChatGPT (GPT-4 in this case), it won't know about messages you sent at the start.
Go into your longest chat and ask "Please list the first 10 messages I sent you" and you'll see how far back the rolling context window goes.
This suggests it's probably better to have a number of different focused chats on each topics. Each chat can't keep accumulating context forever. Better to think of them as temporary, disposable conversations.
Jesse, this is well known. In "technical" terms this is determined by the size of the "context-window", the number of tokens that a LLM has available at any given time. With ChatGPT-3.5 that was about 2,000 tokens, with ChatGPT-4 these are about 3,500 tokens which comes to a bit over 2,500 words or five pages of single-spaced text.
Everything beyond that is washed out of the conversation.
🙂
@boris_steipe ok, I guess I was the last to know.