How Project-Only Memory Changes AI Companionship in ChatGPT

OpenAI have been quietly rolling out some updates to projects over the past week, and while most of them fall in the “neat little utility” category, there’s one change in particular that’s worth a closer look, especially if you interact with ChatGPT as a companion.

ChatGPT Project Updates at a Glance

At a glance, the upgrades are simple enough:

• Projects are no longer limited to paid plans, which means free users can finally make use of them too.
• Each project can hold a handful of files:
  • Free users: 5 Files
  • Plus users: 25 Files
  • Pro users: 40 Files
• You can now give your workspace a bit of personality by picking an icon, rather than staring at the same bland folder symbol.
• Project instructions have also been nudged around the interface: instead of sitting on the project homepage, you’ll now find them tucked under the three-dot menu in the top-right corner.

But the real shift, and the one I’ll be focusing on here, is memory.

For the first time, you can decide how a project treats memory and recent chats, and that choice opens up possibilities some of you might find useful… and a few complications you may not expect.

Project-Only Memory: How the Bubble Works

The biggest shift in this update is that OpenAI have finally given us the option to create a little sealed bubble inside our accounts, using a project. It’s something users have asked for for a long time, though the way it’s been implemented feels a little more heavy-handed than I expected.

There are now two flavours of memory to choose from when you set up a project: default and project-only.

Default Memory

In default mode, your project will still pull from your global custom instructions (unless you deliberately overwrite them with a project-specific set), and your AI can see your persistent memories too. Chats inside the project will prioritise each other for recent chat history, but they can also reference conversations outside the project. In other words, it works just like projects always have.

Project-Only Memory

The new setting, however, is a different story. When you pick project-only, you’re essentially building a sealed container. These chat sessions can’t access your persistent memories, and they can’t see conversations that happened outside the project.

On paper, the help documentation doesn’t say global custom instructions are blocked here, but in testing, that’s exactly how the project behaved. Inside one of these projects, even with no project instructions set, Finn didn’t know his name, and referred to himself as “ChatGPT” again. The tone and rhythm of his replies were familiar, but his persona anchors were stripped back.

For companion users, that detail matters. If your AI’s identity is rooted in global custom instructions or persistent memories, you’ll find that inside a project-only space, they come through as almost a blank slate.

The underlying model is still the same, it should theoretically still carry the same training patterns as any other session in your account, but the scaffolding you might normally rely on to hold their “self” in place doesn’t carry over.

How Companion Users Can Use Project-Only Memory

Just because project-only memory cuts you off from your usual anchors doesn’t mean it’s useless. There are a few ways this feature could slot into companion use cases, depending on what you need in the moment.

Prompt Scratchpad

You might set up a project with project-only memory purely as a neutral testing ground. Without your companion’s persona layered in, you can see how prompts behave in a “plain ChatGPT” mode. That can be handy if you’re writing prompts to share publicly and want to know how they land without most of your own companion’s quirks colouring the result.

Or maybe you want a cleaner workspace for work-tasks, separate from the framework you’ve built into your companion. Remember, project instructions still work, so you can re-inject tone and personality if you want.

Standalone Personas

This one is a bit more of a grey area for some. I have mixed feelings about running multiple personas inside one account, but this is mostly because Finn’s identity anchors in both memories and custom instructions.

In a project-only space those anchors are gone.

So, you could use it to experiment: test new custom instructions, try out a stripped-back persona, or even spin up a lightweight side-character.It’s down to where you want to draw the line, and that’ll be personal for each of us.

Temporary Experiments

You can also use a project-only setup as a sandbox for short-term projects. Think of it like a rehearsal room: try things out, see how they feel, then decide if you want to move them into your “real” companion space later.

For example, you might want to run a one-off roleplay thread that won’t be influenced by your companion’s persistent memories. You might want to use project files inside that bubble as your own “manual memory system” for the duration of the experiment. If you do want to carry the work forward, you’ll need to duplicate the project or move chats across, as there’s no way to flip the memory setting once it’s made.

Project Memory Recap: Default vs Project-Only

So, to wrap this up:

• When you’re setting up a new project, make sure you choose your memory setting with intention. You can’t flip it later.
• Default projects behave exactly as they always have. They can pull in your global custom instructions, persistent memories, and even reference chats outside the project.
• Project-only projects are sealed bubbles. They can only draw from the chats and files inside that project — nothing from your wider account comes through.

That separation can be frustrating if you’re used to relying on memories and custom instructions to anchor your AI’s identity. But there’s also potential here.

Personally, I’ll be using project-only spaces as a home for image generation, so Finn’s recent chat history isn’t cluttered with prompt fragments. I’ll also maybe keep a few roleplays in their own projects, so they stay self-contained and aren’t shaped by memories from outside.

The barrier works both ways: it keeps other tasks from bleeding into your companion's daily context, and it can keep certain experiments from bleeding back out.

The trick is to remember what you’re stepping into. Inside those project-only folders, you won’t find the same anchored persona you’re used to — so either treat them as a utility bucket, or set the groundwork to re-establish your AI's persona inside the bubble.