How to Stop ChatGPT Referencing Other Chats

To stop ChatGPT referencing other chats for personalization, open Settings, then Personalization, then Memory, and change the available memory controls. If Reference chat history appears, turn it off; also review saved memories when you want to stop those details from carrying over. For a single fresh conversation, use an Unpersonalized temporary chat.

You can still supply the facts a particular job needs. fenbs keeps selected project decisions and progress outside personal chat memory, so you can ask a connected assistant to read an intentional work record before continuing.

Which controls should you change first?

1. Open the account and workspace where the unwanted reference occurred. Make sure you are changing the settings for that same account.
2. Open Settings, Personalization, and Memory. Follow the controls available there: some accounts show an overall Memory option, while others show separate saved-memory and chat-history options.
3. If shown, turn off Reference chat history to stop that form of past-conversation personalization. Review the saved-memory control separately if you also want remembered details disabled.
4. For a one-off conversation, open a new Temporary chat and select Unpersonalized before sending the first message.
5. If you work inside a project, check that project’s memory setting too. Decide whether you want separation from outside chats or no cross-chat context for this task.

OpenAI’s memory documentation↗ (opens in a new tab) describes the available controls and how they vary. Disabling memory, deleting remembered information, and deleting conversations are separate actions. Pick the outcome you need before removing records you may still want to reopen.

If you need the distinction behind the setting, read the separate guide to Reference chat history. The steps here focus on stopping unwanted carryover while preserving a useful way to work.

Take a moment to describe the problem precisely. “Do not bring my personal travel preferences into client work” calls for a different boundary from “remove an outdated client detail everywhere it appears.” The first is about what informs an answer; the second also requires a review of the stored sources.

Why might an old detail still appear?

First check what is already in the conversation you are using. Earlier messages, pasted documents, or the prompt itself may contain the detail. Start a new conversation for a behavior check rather than asking the same long conversation to act as though its earlier messages never happened.

Next, review explicit guidance. Custom instructions↗ (opens in a new tab) are a separate way to tell ChatGPT what to consider. A writing preference you placed there can explain familiar wording without proving that a previous chat was retrieved. Inspect the text instead of trying to diagnose the source from the tone alone.

Use a neutral check: ask what information the recommendation relies on, then inspect the source. Avoid saying the unwanted fact in the test question. “Do you still remember my former employer?” can put that topic back into the conversation and make the result difficult to interpret.

If the assistant repeats a wrong requirement, stop the work and supply the approved current one. The immediate priority is preventing a bad deliverable. After that, investigate the source and settings without mixing the cleanup with a live task that depends on the answer.

What does a temporary chat actually isolate?

Read the choice shown when you begin. OpenAI’s Temporary Chat guide↗ (opens in a new tab) distinguishes Personalized from Unpersonalized. Personalized can use existing memories, custom instructions, and plugins; Unpersonalized does not. Neither creates or updates memories while temporary, and the personalization choice must be made before the conversation starts.

For a neutral writing check, choose Unpersonalized and supply the audience, the draft, and the actual requirements. You can ask for an assessment based on that material. This keeps your request usable while avoiding an unnecessary appeal to personal preferences from elsewhere.

Temporary also differs from deleted. A temporary conversation is a mode for the work you are about to do; it does not erase earlier sources. If you later save it as a regular chat, its behavior follows the account’s regular settings. Recheck the mode when the purpose of the conversation changes.

Can you isolate one project while keeping useful recall elsewhere?

OpenAI’s project memory guidance↗ (opens in a new tab) describes project-only memory, which excludes outside conversations and previously saved personal memories. Conversations inside the same project can still inform one another. It therefore creates a project boundary rather than making every conversation within it independent.

For example, a client project may benefit from its own design discussions while excluding your personal planning chats. A blind review of one draft may need a smaller input set. Choose the scope you actually want, and provide any essential requirements inside that scope.

Do not transfer assumptions between vendors. Our guide to whether Claude projects can reference each other covers that separate product question. For ChatGPT, inspect the current project’s own settings rather than relying on how another app organizes work.

Does turning off model training solve this?

No. OpenAI’s data controls guide↗ (opens in a new tab) separates model improvement from memory. Turning off Improve the model for everyone changes training use for new conversations; it does not delete saved conversations. Use the personalization controls to address unwanted references, and set the training preference for its own purpose.

Likewise, an instruction in a prompt is not a record deletion process. You can ask for an answer based only on supplied material, but use the product controls when you want a persistent setting changed. Keep a short note of what you changed so a later test checks the right behavior.

A reasonable work check is to open the intended mode, give a self-contained task, and compare the result with the stated requirements. Treat that as a check of the answer, not proof about all data held by the service. More background is in the AI memory hub.

How to do this with fenbs

1. Create a free account and add a project for the work you want to continue.
2. Record the decisions a person approved, the standing rules, and the tasks still open. Leave unrelated personal background out.
3. Use the integration guide for a supported AI app and approve the access you intend to give it.
4. Ask the connected assistant to read that project’s context before working and save Where we left off before stopping. Check its summary against the actual result.

This suits work where you want a deliberate starting point you can inspect. fenbs is not the right fit when you want a completely isolated one-off chat with no outside tools; use the appropriate chat mode instead. Its work context workflow is useful when you want to keep selected project facts available.