If memory is turned off it might have to do with OpenAI experimenting with extended memory functionality over previous chats or you want to like this bug report:
As noted memory was off. When asked to recall a prior hypothesis I had, which GPT had named ‘Kernel Theory’, it did so. It did this at the start of a new, otherwise blank conversation. The interaction is copied below.
I returned to this topic, after learning this should have not been possible. GPT stated that its answer was derived from the name alone, which is completely impossible given the level of detail (again see below).
It then appeared to lie, repeatedly, until I quoted its own answer,…