This was a controlled test involving GPT-5 and GPT-4o, which are supposed to have completely separate memory contexts per OpenAI’s documentation.
I typed a unique phrase into GPT-5, then deleted the chat.
I opened a new GPT-4o chat, with no shared history, and asked a vague follow-up.
GPT-4o quoted the exact phrase back.
In another test, it referenced material from an entirely different GPT-5 chat, never typed into 4o at all.
So this isn’t memory acting strange—it’s GPT-4o accessing GPT-5-only content, even from deleted sessions.
According to OpenAI:
“Chats with GPT-4o don’t currently use memory” and “Memory is unique to each model.”
If this is replicable, it’s not just a quirk—it’s a model boundary violation.
Happy to clarify details if others want to try reproducing.