You keep a project for a client, with project instructions and their contracts uploaded. Memory is on. In a plain chat outside the project you ask a question about that client, and the answer is vaguer than you expected. Which explanation fits what each feature actually does?

The principle — three different mechanisms supply context in ChatGPT, and they have three different scopes. Reasoning about a surprising answer means asking which of them could have reached the chat you were in.

Why the key is correct — a project is a workspace holding grouped chats, uploaded reference files and its own instructions, and OpenAI states that project instructions apply inside the respective project only. The contracts are files in that workspace, so a chat started outside it has no access to them. Memory is the one mechanism that is not scoped this way: it remembers useful context from your chats to personalise your experience, so general knowledge about the client can plausibly travel with you. That combination is precisely a vaguer answer rather than no answer at all, which is what you observed.

Why the others are wrong — a weaker reply does not prove memory captured nothing, and the reasoning that says so confuses two kinds of context. Project instructions are not promoted to account level by being written; their whole purpose is to override the global ones inside one body of work. And history is a list you can reopen rather than context supplied to new conversations, so it could not have carried the files even in principle.

Remember this — the diagnostic question is always where the information physically lives. Files live in the project or the chat you attached them to. Instructions live at whichever scope you set them. Memory is derived context about you, which is why it is the only one that follows you around, and why the memory summary is where you check what it actually holds.

Sources — OpenAI's Projects in ChatGPT article and Memory FAQ.