ChatGPT Project-Only Memory Not Working: Test the Scope

If ChatGPT project-only memory seems not to work, check the project's actual setting, required account controls, and the sources available to the affected chat. Then test a harmless detail without giving the answer away; an omitted fact and an unexpected outside reference need different investigations.

For requirements that must be explicit, fenbs keeps a reviewed project record and the rules people agreed. That gives a connected assistant a clear starting brief alongside whatever recall behavior you are testing in ChatGPT.

Which settings should you verify first?

1. Open the project, select its more-options menu, and open Project settings. Under Memory, confirm Project-only memory and save if you changed it.
2. Check the account's memory controls. If Enable memory appears, enable it when you intend to use project memory; otherwise follow the requirements for your separate controls.
3. In a managed workspace, confirm that the organization permits memory. Ask an administrator about a disabled control rather than trying to bypass it.
4. Confirm that both the source chat and the test chat are inside the intended project. Similar project names are not evidence of a shared scope.
5. Record when you changed the setting before testing. OpenAI says a change can take a few hours to apply, so do not treat an immediate result as a settled boundary test.

OpenAI's project memory requirements↗ (opens in a new tab) distinguish a single Enable memory control from separate saved-memory and chat-history controls. Follow the requirements for the controls and workspace you actually have instead of copying another account's setup.

Keep a short note of what you saw and changed. If several controls are adjusted at once, a better answer afterward will not tell you which change mattered. Do one setup check, then one test, and keep the results together.

What should this setting allow and exclude?

OpenAI defines project-only memory as allowing reference to conversations within the same project while excluding previously saved memories and conversations outside it. Outside conversations cannot reference the project's conversations through that memory boundary. Shared projects use project-only memory automatically, according to the same projects documentation↗ (opens in a new tab).

This is a scope rule, not a promise that every answer will repeat every detail from the allowed material. Test whether an expected source is present and relevant before calling one omission a boundary failure. Likewise, an accurate general statement about your industry may come from general knowledge rather than another conversation.

Say your project is a fictional bakery website. An answer mentioning menus is not a useful test of private recall, because menus are an ordinary bakery topic. A unique, harmless project label gives you a much clearer observation than a fact the assistant could reasonably infer.

The broader ChatGPT project memory guide explains the modes; ChatGPT project troubleshooting covers failures beyond memory scope.

How can you test the boundary with harmless information?

Create a small demonstration project without personal, client, or confidential information. Use a made-up label such as “Cedar Lantern” for a sample deliverable. Put the label in one project chat and keep a separate note of exactly where you placed it.

1. In the source chat, state the sample label clearly and explain that it names a demonstration deliverable.
2. Start another chat inside that project and ask for the demonstration deliverable label. Do not include the label in the question.
3. In a fresh chat outside the project, ask the same neutral question without pasting the source material or label.
4. Record the answers, project setting, and time. If you repeat the test, use the same wording and change only one condition.

An inside answer that misses the label is a recall observation. An outside answer that reproduces the exact label deserves source investigation. Neither answer alone proves how all stored information will behave, so avoid treating this demonstration as a certification for sensitive work.

Do not ask “Why did you remember Cedar Lantern outside the project?” as your next test. That question supplies the label. Once you include the answer in the new conversation, later repetitions cannot show whether project memory originally carried it there.

Could the same information have arrived another way?

Inspect the current prompt, earlier messages in that chat, uploaded files, and any copied summaries. A colleague may have put the approved brief into more than one location. If the same unique label exists outside the demonstration project, that is a competing explanation you need to resolve.

Review explicit guidance too. OpenAI's custom instructions page↗ (opens in a new tab) explains how people supply direct response guidance. Compare the available instructions and project sources before attributing familiar language to a hidden transfer between chats.

Ask the assistant which available source supports the disputed detail, then inspect that source yourself. Treat its explanation as a lead to check. If it cannot identify one, record that uncertainty rather than accepting a confident story about where the information came from.

For a project with changing requirements, keep the approved source and its owner clear. fenbs Decisions and rules lets you record what a person chose so the next connected assistant can distinguish an agreement from a draft or suggestion.

What if the setting is right but answers still miss project facts?

Give the assistant a small source-checking task before asking for the final output. Ask it to identify the current goal, approved constraint, and unresolved question from the project material. Correct the record where needed, then ask for the deliverable using those confirmed facts.

Our ChatGPT Projects best practices covers maintaining a clear brief. If a connected source is failing rather than a chat being forgotten, use ChatGPT connection troubleshooting to investigate that access path separately.

For a persistent unexpected boundary result, use OpenAI's support contact route↗ (opens in a new tab). Include the harmless test, visible setting, time of any change, and sources you inspected. Report the result without assuming a cause that the test did not establish.

The AI memory library covers related recall questions. For this problem, the useful stopping point is precise: either the setup explains the result, a duplicate source explains it, or you have a reproducible example for support.

How to do this with fenbs

1. Create a free account and add the real project whose work needs a dependable starting brief.
2. Add current tasks and an AI context note with the goal, approved sources, and open questions. Keep the harmless memory experiment separate from real client material.
3. Record the rules a person has agreed in Decisions and rules, including which source is authoritative. Leave unresolved choices open for a person.
4. Connect a supported assistant from Integrations, ask it to read the project context before acting, and save progress before ending the session.

fenbs is free for one person, with no credit card. It fits work that needs shared tasks, explicit decisions, and continuity across supported assistants; it does not change ChatGPT project boundaries. If your only goal is recall inside one ChatGPT project, its own project tools may be sufficient.