ChatGPT Projects Can Now Change Their Memory Boundaries, And That Matters for Governed AI

By Wendy Keir | EmpowerAi™

OpenAI has made what looks, at first, like quite a small change to ChatGPT Projects.

You can now take an existing Project and change its memory setting.

Until this update, the memory choice was something you made when you created the Project. If you later realised that the work really needed to stay inside its own contextual boundary, you could not simply change it. With the update released on 14 August, eligible unshared Projects can now move between Default memory and Project-only memory through Project settings. Shared Projects remain Project-only.

I think this is more important than the setting itself makes it sound, because it touches something we are going to have to become much better at as AI starts sitting inside more of our work.

We need to decide what the AI is allowed to carry from one environment into another.

Memory is useful until the wrong memory appears

One of the reasons ChatGPT has become considerably more useful over the past couple of years is that conversations no longer have to begin from zero every time.

The system can understand preferences, ongoing work and information from previous conversations, depending on the settings being used. That continuity can save a great deal of repetition.

There is also a point where continuity stops being helpful.

Imagine that you are an author working on two books with quite different methodologies. You might also have client conversations, business strategy work, personal discussions and hundreds of miscellaneous chats sitting elsewhere in ChatGPT.

You probably do not want all of that context influencing a conversation about one particular manuscript.

The same problem appears inside an expert-led business. A consultant may be working with several methodologies. A coach might have different programmes for different groups of clients. Somebody creating intellectual property may have early ideas that were later abandoned sitting alongside the version that eventually became their recognised framework.

More memory is not automatically better memory.

What matters is whether the context belongs.

Project-only memory creates a clearer boundary

With Project-only memory enabled, ChatGPT can use conversations from inside that Project to provide context, while conversations and memories from elsewhere are kept out. Information from the Project is also prevented from becoming part of the memory used in unrelated chats.

That creates a much clearer container.

If I had a Project for a particular book, for example, the conversations taking place while I worked on that book could inform one another. The work could develop over time without every new conversation requiring a complete explanation of what had already happened.

At the same time, my completely unrelated conversations elsewhere in ChatGPT would not quietly become part of that environment.

There is something quite important in that distinction.

A useful AI system needs context.

A governed AI system also needs boundaries around that context.

This becomes more significant when the knowledge is intellectual property

The further AI moves into expert businesses, the more often it is going to encounter knowledge that has structure and ownership.

A methodology is not simply a collection of useful information.

It has relationships between ideas. It may contain terminology that means something very specific inside the author's work. There may be deliberate exclusions, ethical boundaries and decisions the expert has made about what they do and do not recommend.

If an AI system begins importing information from elsewhere simply because it happens to remember it, the conversation can gradually move away from the original methodology without anybody noticing.

This is one of the reasons I think the conversation about AI memory needs to become more sophisticated.

We have spent a lot of time asking whether AI can remember us.

The more useful question for expert work may be whether it knows what it should remember here.

Those are not quite the same thing.

This is very close to a question I work with when building AI Book Companion™

When I build an AI Book Companion™, I have to decide what belongs inside the Companion's knowledge environment.

The obvious source is the book.

There may also be worksheets, supporting explanations, frameworks or other material the author has explicitly approved.

Then there is everything else the underlying AI may know.

That is where governance becomes important.

A reader may ask a perfectly reasonable question about something the author has never discussed. The underlying model may be capable of answering it from its wider knowledge, but doing so can blur the distinction between the author's methodology and general AI-generated advice.

The fact that the model knows something does not necessarily mean the Companion should use it.

Project-only memory is obviously a different feature and OpenAI has not presented it as a Book Companion architecture. What interests me is the principle underneath it.

Context is becoming something we can increasingly contain rather than simply accumulate.

That is a useful direction.

Authors may need to start thinking in knowledge environments

For a long time, authors have organised their work around documents.

There is a manuscript.

There are notes.

There may be a research folder and perhaps some worksheets or supporting resources.

AI introduces another layer because those documents can now sit inside a conversational environment.

The reader, writer or expert can ask questions across them.

Ideas can be compared.

Previous conversations can influence later ones.

Once that happens, the structure surrounding the documents starts to matter almost as much as the documents themselves.

An author might eventually have one environment for developing the manuscript, another for marketing the book and another governing the reader-facing Companion.

Those environments may contain some of the same material.

They do not necessarily need the same memory.

The ability to change the boundary matters as projects evolve

There is another practical part of this update that I quite like.

You do not always know how important a piece of work will become when you begin it.

A Project might start as a convenient place to collect a few conversations. Six months later it may contain an entire methodology, a book manuscript and a significant amount of intellectual property.

Being able to change its memory architecture without abandoning the existing Project means the governance can evolve with the work.

That feels much more realistic than expecting people to understand every boundary before they begin.

Good AI governance is probably going to work like that in practice.

We start using something.

We understand where the risks and opportunities actually sit.

Then we tighten or change the environment as the work becomes more important.

Memory is becoming part of AI architecture

I suspect we are going to talk much more about this over the next few years.

Memory used to feel like a feature.

ChatGPT remembered something about you and the next conversation became slightly easier.

As AI takes on longer pieces of work, memory begins to become part of the architecture underneath the system.

Which conversations belong together?

Which knowledge should remain isolated?

What is allowed to move from one environment to another?

Who gets to decide?

Those questions become particularly important when the AI is representing someone's expertise rather than simply helping them write an email.

OpenAI's latest Project change does not answer all of them.

It does give users another piece of control.

For authors and expert-led businesses building knowledge-based AI systems, I think that is worth paying attention to.

Sources

OpenAI — ChatGPT Release Notes: ChatGPT app experience updates, 14 August 2026.