By skyforbes However, today I sent a single phrase for a title I wanted to post here on Reddit: "This fragment of a flyer on a street pole looks like a tiny bored creature"
Its response was the same phrase, but in the list of corrections it said:
Original text: It until episode 4
Corrected to: until (Removed the extra word to fix the phrasing)
This was odd. I asked to give me the entire prompt it received. It did and I recognized it: it was a text I sent a couple of days earlier in the same project.
I immediately understood it was referencing past conversations, and I thought that the setting to reference chat history had switched from off, my preference, to on again. But the setting is still off. It shouldn't reference past chat, but this was clear evidence that even with the setting turned off, chats are not totally isolated.
Does anyone know about this limitation? Although it is very practical to use past chats to have persistence of knowledge, what I found is that context I didn't want for certain conversations would leak and interfere with the responses.