How to Stop ChatGPT from Using Old Information in New Chats

If ChatGPT keeps bringing old preferences, outdated facts, or instructions into a new chat, starting another standard chat may not be enough. The information can come from saved memory, referenced chat history, custom instructions, the current project, or the conversation you are still using.

The fastest clean test is to open a Temporary Chat. Temporary Chat does not use or create memories for personalization, but it still follows enabled custom instructions. If the unwanted information appears there too, inspect Custom Instructions and any instructions or files attached to the current GPT or workspace. For a permanent cleanup, remove the relevant memory and also delete the original chat when the information must be fully removed.

Identify Where the Old Information Is Coming From

Before deleting anything, determine which context source is responsible. These sources behave differently:

• Current conversation: Messages and files already in the open chat remain part of that conversation. A follow-up prompt does not create a clean context.
• Memory: Depending on your account and rollout, Settings may show a Memory summary, saved memories, Reference chat history, or a combination of these controls.
• Custom Instructions: Account-level instructions can affect new and existing chats, including Temporary Chats.
• Project context: Chats inside a project can use project instructions, files, and other conversations allowed by that project's memory mode.
• Custom GPT or connected tool: A GPT can have its own instructions, knowledge files, or actions that influence the response.

A browser refresh, cache clear, or app reinstall does not remove account-level memory. Those actions may fix a display problem, but they do not replace the memory and personalization controls.

Use Temporary Chat for an Immediate Clean Test

Open a new chat and select the Temporary control near the top of the ChatGPT interface. OpenAI states that Temporary Chats do not appear in history, do not create memories, and are not used to improve its models. A copy may still be retained for up to 30 days for safety purposes.

Ask a neutral question that previously triggered the unwanted information. Do not include the old detail in the test prompt. For example, if ChatGPT keeps assuming an outdated job role, ask:

What background information are you using to tailor this answer?

If the outdated detail disappears, memory or chat history was probably involved. If it remains, check Custom Instructions, the selected GPT, project instructions, or details already included in your prompt.

Temporary Chat is not a complete privacy guarantee. It still follows enabled custom instructions. OpenAI also states that limited safety-relevant context from previous conversations may be used in rare, high-risk situations to improve safety. Information sent through a GPT action or another connected service can be handled under that third party's policy.

Review What ChatGPT Remembers

Open your profile menu, then go to Settings > Personalization > Memory. The exact layout can vary because OpenAI is rolling out an updated memory experience. You may see a memory summary, a Manage memories screen, separate saved-memory and chat-history controls, or fewer options depending on the plan and workspace.

You can also ask ChatGPT:

What do you remember about me that could affect your answers?

Treat the response as a review aid rather than a complete audit. OpenAI notes that a memory summary may not display every detail that can be drawn from prior chats.

Delete or Correct the Relevant Memory

If the old detail appears in Manage memories or the memory summary, remove it there or ask ChatGPT to forget it. When the information is merely outdated, replace it with the correct current detail rather than clearing unrelated memories.

Turning memory off is not the same as deleting what was already remembered. OpenAI's guidance says that disabling saved memory does not remove existing saved memories. To stop a specific item from returning, delete or update that item directly.

After the change, start a new standard chat and ask a neutral question again. OpenAI notes that deleted memories may take a few days to stop being referenced in every chat, so avoid repeatedly adding and deleting the same item while testing.

Delete the Source Chat When Full Removal Matters

Saved memories are stored separately from chat history. Deleting a chat alone therefore may not remove a memory created from that conversation, and deleting a memory does not erase the original messages.

For full removal of a sensitive or incorrect detail:

1. Delete the related saved memory or update the memory summary.
2. Find the chat where the information was originally shared.
3. Export or copy anything you still need.
4. Delete the chat from its three-dot menu.
5. Confirm the warning only after checking that the conversation is no longer needed.

Deleted chats cannot be recovered through the interface or support. OpenAI says they are removed from the account view immediately and scheduled for permanent deletion within 30 days, subject to stated legal, security, or de-identification exceptions.

Archiving is safer when you only want to reduce sidebar clutter. It keeps the chat in the account and is not a substitute for deleting sensitive material or clearing a saved memory.

Turn Off Reference Chat History When Available

Some accounts expose separate controls for saved memories and Reference chat history. Chat-history reference can use useful information from previous conversations even when it was not stored as a visible saved-memory item.

If you want saved preferences to remain but do not want prior chats influencing new work, leave saved memories on and turn Reference chat history off when your interface provides that choice. OpenAI states that turning this setting off also deletes information remembered from past chats from its systems within 30 days.

If your account uses the newer combined Memory experience and does not show the separate toggle, use the available Memory controls, Temporary Chat, or a project with project-only memory. Do not rely on instructions for a menu that your current rollout does not display.

Check Custom Instructions Separately

Open Settings > Personalization > Custom Instructions or the equivalent Customize ChatGPT section on mobile. Look for old job titles, formatting rules, preferred tools, locations, names, or assumptions that no longer apply.

Custom Instructions are separate from memory. OpenAI says changes apply immediately across chats, and Temporary Chat still follows enabled custom instructions. Disable customization for a clean test, or edit only the outdated line so useful preferences remain.

For reusable instructions that should apply only to one workflow, keep them in a project or in the opening prompt instead of placing them in account-wide Custom Instructions. The guide How to Write Reusable AI Instructions for Consistent Results explains how to separate stable rules from task-specific inputs.

Use Project-Only Memory for Work That Must Stay Separated

For long-running work, create a new project and choose project-only memory when that option is available. OpenAI says this mode prevents previously saved personal memories from being referenced and blocks chats outside the project from influencing chats inside it. Conversations within the same project can still reference one another.

Project-only memory must be chosen when creating a new project; an existing default-memory project cannot simply be converted. Availability also depends on account and workspace settings. Review project instructions and uploaded files because they remain intentional context inside the project.

If one conversation inside a project contains incorrect information, move or delete that conversation rather than expecting a new chat in the same project to ignore it automatically.

Do Not Confuse Memory with Model Training

Improve the model for everyone is a Data Controls setting about whether conversations may be used to improve models. It is not the switch that determines whether ChatGPT personalizes answers from memory.

Turning model improvement off does not automatically clear memory, and turning memory off does not necessarily change the model-improvement setting. Review both sections when privacy is the concern:

• Settings > Personalization: memory, custom instructions, personality, and related personalization controls.
• Settings > Data Controls: model-improvement preference, exports, archived chats, and account data controls.

Verify the Cleanup Without Reintroducing the Detail

Use a controlled test after making changes:

1. Write down the outdated assumption outside ChatGPT so you do not need to paste it into the test.
2. Start a Temporary Chat and ask a neutral question.
3. Start a normal new chat and ask the same question.
4. Compare the two answers for the unwanted assumption.
5. Check Custom Instructions or project context if both answers still contain it.

Do not test by repeatedly telling ChatGPT the old information. Doing so can place it back into the current conversation and make the result difficult to interpret.

Privacy and Managed-Account Limits

Do not place passwords, authentication codes, customer records, medical details, confidential contracts, or unredacted company data into memory or Custom Instructions. Remove direct identifiers and use test data when possible.

Work and school administrators may control memory, projects, retention, connected apps, or data settings. If a control is missing or locked, follow the workspace policy rather than trying to bypass it. For source-based work, How to Use ChatGPT Study Mode with Your Notes and PDFs includes checks for private files and source verification.

What Not to Do

• Do not assume that opening a normal new chat creates a blank slate while memory is enabled.
• Do not delete all chats before reviewing which information and conversations you still need.
• Do not treat archiving as deletion of saved memory.
• Do not clear browser data expecting account-level memory to disappear.
• Do not trust a corrected answer without checking important facts against the current source.

When old information changes the substance of an answer, verify the final result rather than assuming the cleanup fixed every claim. A Verification Prompt That Helps Reduce AI Hallucinations provides a source and uncertainty checklist for high-impact outputs.

Official OpenAI References

• Memory FAQ
• How Reference Saved Memories Works
• Temporary Chat FAQ
• ChatGPT Custom Instructions
• Projects in ChatGPT
• How to Delete and Archive Chats in ChatGPT
• Data Controls FAQ

Use the Smallest Control That Solves the Problem

Use Temporary Chat when you need one clean session. Delete or correct one memory when a specific fact is wrong. Turn off chat-history reference when past conversations should not shape new work. Edit Custom Instructions when the same outdated rule appears everywhere. Create a project with project-only memory when a continuing workflow needs a durable boundary.

This approach removes the unwanted context without destroying useful history or disabling every personalization feature.