Hi everyone,
I’m running an experiment that rates a number of research papers with ChatGPT’s Deep Research tool and then has o4-mini assign scores. To check reliability, I want to repeat the whole workflow several times independently.
OpenAI’s release notes say ChatGPT now keeps a cross-chat “memory” unless I turn it off or use Temporary Chat.
The FAQ confirms Temporary Chat blocks that memory.
Meanwhile, the docs state the Chat Completions API is stateless, each request sees only the messages I send.
My questions
- If I simply click “New Chat” (with memory on), could ChatGPT still recall information from earlier runs and skew the ratings?
- Does starting every replicate in Temporary Chat guarantee isolation?
- Am I correct that the safest approach is to script this in the API, since it never retains prior context?
- Any best practices for seeding or temperature settings to keep runs comparable?
Appreciate any insights or pointers to official docs, thanks!