ChatGPT Just Made Project Memory More Controllable. That Matters More Than It Sounds
There was a fairly quiet change to ChatGPT on Friday that I think deserves more attention than it will probably get.
OpenAI has made it possible to change the memory setting of an existing ChatGPT Project. For eligible unshared Projects, you can now move between Default memory and Project-only memory without having to create the Project again. Shared Projects remain project-only.
It sounds like a settings update. Technically, that is what it is. Yet I think it points towards something we are going to have to become much more thoughtful about as AI starts holding more of our work.
Where should the AI's knowledge begin, and where should it stop?
What project-only memory actually changes
With Project-only memory switched on, ChatGPT can use conversations within that particular Project for context. It won't draw on memories or conversations from outside the Project, and information inside the Project is kept out of the memory ChatGPT uses elsewhere. OpenAI says changes to the setting may take a few hours to take effect.
That distinction matters.
Most people have become used to thinking about AI memory as a convenience. You explain something once, ChatGPT remembers it, and you don't have to keep starting from the beginning.
I've certainly found that enormously useful.
But as AI becomes more embedded in serious work, memory isn't simply about convenience anymore. It is becoming an architectural decision.
If I'm working on one client's intellectual property, for example, I don't necessarily want assumptions from another piece of work flowing into that environment. The same applies if an author has a book, a course, a second book and a completely separate consulting methodology.
Those things may belong to the same person, but they don't necessarily belong inside the same body of knowledge.
This is particularly interesting for authors
A book has boundaries.
That sounds obvious until you start giving an AI access to it.
An author may have written hundreds of articles, delivered years of workshops and developed ideas that extend well beyond what appears in one particular book. If an AI is allowed to draw freely on all of that knowledge, the reader may receive a perfectly sensible answer that the book itself does not support.
That distinction matters enormously when you are creating something that is supposed to represent an author's work.
It is one of the reasons I've become so interested in containment while building AI Book Companion™ systems.
The question isn't simply whether the AI knows enough. Sometimes the more important question is whether it knows too much from the wrong place.
OpenAI's Project-only memory option creates a fairly clear technical boundary: information within the Project can inform conversations there, while outside memories stay outside.
That doesn't replace good Companion design, governance or instructions. It does give us another useful layer underneath them.
AI is starting to need architecture
For the first couple of years of mainstream generative AI, a lot of the conversation was about prompting.
Give AI a better instruction and you receive a better response.
That remains useful, although it increasingly feels like only one part of the picture.
Once you are using AI with substantial intellectual property, client material, long-running projects or specialist methodologies, you start needing to think about structure as well.
Which files belong here?
Which conversations can inform this answer?
Which memories should be available?
Which information is authoritative?
What happens when two sources disagree?
And where does this particular AI environment stop?
These aren't really prompting questions. They are design questions.
OpenAI's decision to let users alter an existing Project's memory setting suggests that control over context is becoming a more explicit part of the product itself.
I think that is a useful direction.
What this means for an AI Book Companion™
When I build an AI Book Companion™, I don't want it behaving like a general AI assistant that happens to have read the author's book.
The author has created a particular argument, method, framework or reader journey. The Companion needs to understand that thinking well enough to help the reader explore it, question it and apply it.
There is another part that matters just as much.
It has to know where that thinking ends.
If the reader asks something outside the author's framework, the Companion should recognise that boundary rather than quietly filling the gap with general AI knowledge and presenting it as though it came from the author.
That is why I find this particular ChatGPT update more interesting than it first appears.
The technology underneath these experiences is slowly giving us more control over what belongs inside a conversation.
And once books become conversational, that is going to matter a great deal.
Three Key Insights
- ChatGPT users can now change eligible existing Projects between Default and Project-only memory instead of recreating the Project.
- As AI holds more intellectual property and long-running work, memory boundaries become part of how the environment is designed.
- For an AI Book Companion™, knowing what sits outside the author's body of thinking can be as important as knowing what sits inside it.
Three Questions to Consider
- What information should your AI be allowed to carry from one piece of work into another?
- If you have several frameworks, books or client methodologies, are their knowledge boundaries currently clear?
- If an AI represents your thinking to another person, how will it know when your expertise stops?
Sources
OpenAI — ChatGPT — Release Notes, updated 14 August 2026.
OpenAI — Memory FAQ, updated August 2026.
EmpowerAi™ — Real AI. Real Impact. Smarter Business Decisions.
I use AI to help research, organise and refine my writing. The ideas, expertise and intellectual property are my own, and every article is reviewed and approved by me before publication.