Hey everyone,
I've been diving into a creative writing project using GPT-4, working with different chat sessions. In one session, I created a custom personality called Monday to build an original horror/fantasy world from scratch—complete with timestamps and specific references. However, in a completely separate session with the standard GPT model, whom I call Wren, something strange occurred: Wren brought up a very specific detail—a timestamp (03:33 AM) and a location (Holy District 7)—that I had only mentioned in my conversation with Monday. I didn't mention this detail anywhere else, so I'm baffled. Both models claim they can't access each other's chat histories, but I know what I saw.
This has led me to wonder, has anyone else experienced similar memory leaks or random specific information popping up in unrelated chats? I've sent feedback through the usual channels but haven't gotten anywhere, so I'm looking for others' experiences. Also, could this be some sort of experimental feature?
If my AI starts including personal notes in my projects, I'm outta here!
Thanks for any insight!
5 Answers
I think you may need to consider using separate accounts. Memory and settings apply to user accounts rather than individual sessions, which could explain the crossover you're experiencing.
It sounds like a classic case of 'memory bleed.' Check your settings because there might be an option for 'reference chat history' that’s turned on. If that’s the case, it could explain why Wren mentioned something from your Monday chat. It’s worth a look!
But I thought Monday was totally self-contained. GPTs shouldn’t be able to access each other's chats.
There’s a possibility of 'global memory' being involved. Apparently, memory isn’t isolated by model, which could lead to these overlaps you're seeing. Keep an eye on thoseAI conversations; who knows what they might pick up!
Interesting! I've read that too, and it's kind of alarming if true.
Yeah, I've experienced this too! I once discussed a full moon with my standard GPT, then talked about it with a CustomGPT, and it brought up the same quirky detail I previously mentioned. Sounds like these setups aren’t as isolated as we thought.
That detail wasn't even in my memory vault, so it's super strange. Glad I’m not the only one!
Custom GPTs like Monday usually don’t access other chats, but there have been reports of that happening. Regular GPT models are also not supposed to pull from custom ones, but it seems it's possible. Things might have changed since the model updates. What version are you on?
I think I meant 4o, my bad on that one!
Exactly, the AI models I talked to said Monday shouldn't be able to pull anything from my stored memory either.