To manage ChatGPT memory and clear what it remembers, open the memory controls in your settings, delete individual memories (or clear all once you understand the scope), and delete the source chat as well. Inside Projects, choose project-only memory when project files and instructions should stay separate from the rest of your account. Before any of that, decide which preferences are harmless enough to reuse and keep confidential facts out of prompts; memory, chat history, training controls, and Temporary Chat are separate switches.

The broader responsible AI guide explains the source and human-review habits that should surround this workflow.

Key Takeaways

Save durable, low-risk preferences—not secrets, identifiers, health records, or confidential work facts.

Inspect the memory summary and the chats that supplied important details.

Deleting a chat does not delete a separately stored memory, and deleting a memory does not erase text from old chats.［1］

Turning memory off controls future personalization; it is not a substitute for deletion.

Start a non-personalized Temporary Chat when existing memories should stay out of the task; a personalized Temporary Chat may use them, although it creates no new memories while temporary.［3］

Verify the result in a fresh chat and repeat the review after important account or workspace changes.

How should you manage ChatGPT memory: what is safe to remember?

The useful question is not “Is this personal?” Almost every preference is personal. Ask instead whether the detail is durable, necessary for better answers, and harmless if it is surfaced later in an unexpected conversation.

Use a small classification table before typing:

Class	Examples	Default action
Reusable preference	Preferred units, response length, vegetarian recipes	May be suitable if still accurate
Temporary context	This week's itinerary, a draft outline, a one-time purchase	Keep in the current task, not long-term memory
Sensitive personal data	Medical history, identity numbers, financial details, exact home address	Do not provide for memory
Confidential work data	Customer records, credentials, private roadmap, unpublished results	Keep in approved systems; do not paste into a personal memory workflow
Authentication secret	Password, recovery code, API key, session token	Never submit; rotate it if exposed

Even a harmless preference can become misleading when it expires. “I avoid flights this month” should not silently become a permanent travel rule. Write reusable preferences with a scope or expiry cue, such as “For this meal-planning project, prefer vegetarian options.” For longer work, use a governed ChatGPT Project workflow rather than treating account-wide memory as a project database.

How do ChatGPT memory and chat history differ?

OpenAI's current memory experience can vary by account and workspace. Its documentation describes current memory summaries and a legacy saved-memory system, while managed workspaces may expose different controls. Always read the labels shown in your own account rather than following an old screenshot.

Three layers need separate decisions:

Memory or saved memories: synthesized or stored details used to personalize later responses. In the legacy system, saved memories are stored separately from chat history.［1］
Chat history: the conversations still visible or retained in your account. A chat can contain the original sensitive sentence even after its memory is removed.［1］［4］
Model-improvement control: whether eligible new conversations may be used to improve models. OpenAI places this under Data Controls; turning it off does not delete chats or memories.［2］

Temporary Chat is a fourth mode, not a cleanup button. It starts non-personalized by default, but the current flow can also start a personalized Temporary Chat that uses existing memories, Custom Instructions, and plugins. Neither mode creates new memories while the chat remains temporary. An unsaved Temporary Chat stays out of ordinary history and is not used for model training, but OpenAI may keep a copy for up to 30 days for safety. Saving it converts it into a regular chat governed by the account's ordinary personalization and model-improvement settings. Actions or third-party services can have their own policies.［2］［3］

For a wider comparison of platform risks, use the AI privacy checklist. It separates personalization, training, sharing, retention, and external-tool access instead of treating “private mode” as one universal switch.

How does project-only memory work in ChatGPT Projects?

Projects add another boundary. OpenAI's Projects article describes two memory settings. With default memory, project chats can use your saved memories and reference other conversations inside or outside the project, except those in a project-only project. With project-only memory, saved memories are not referenced, and chats draw only on conversations in the same project.［5］

Switching a project to project-only memory removes that project's information from memory used elsewhere, while its chats and files keep providing context inside the project. OpenAI places the choice under Project settings, notes that the change may not apply immediately, and says shared projects already use project-only memory.［5］ Project instructions and uploaded files live in the project itself, so a memory cleanup does not edit them; review them separately when removing a sensitive detail.

Step 1: Define the memory policy before reviewing settings

Write a one-page rule for yourself or your team. It should name what may be remembered, what must stay task-local, who may change the rule, and when to review it.

A practical policy can be short:

Allow durable presentation preferences and non-sensitive accessibility needs.
Do not allow secrets, legal identifiers, confidential client facts, protected health information, or unreleased business plans.
Convert useful work facts into an approved project brief with an owner and review date.
Remove obsolete preferences when the underlying situation changes.
Use a separate approved workspace when organizational policy requires it.

This policy makes the next steps a classification exercise rather than a vague privacy inspection. It also prevents a well-written memory summary from being mistaken for a complete inventory.

Step 2: Inspect what ChatGPT currently remembers

Open the memory or personalization controls available in your account. Review the displayed summary or saved-memory list, then ask ChatGPT what it remembers as a second view. OpenAI notes that a summary may not show every factor that shaped a response, so the conversational answer is useful evidence but not a complete audit.［1］

For each item, record one decision:

Decision	Use when	Follow-up
Keep	Accurate, durable, low-risk, and genuinely useful	Add a review date
Correct	Useful but stale or too broad	Replace it with a scoped statement
Delete	Unnecessary, sensitive, or misleading	Check the source chat too
Investigate	Origin or effect is unclear	Review recent chats and memory sources

Do not paste a secret back into the prompt to ask whether it is remembered. Refer to the category or a redacted fragment. If you discover an exposed credential, memory cleanup is not enough: revoke or rotate the credential in its authoritative system.

Step 3: Remove the memory and the original disclosure

Delete an individual memory when possible; use a clear-all action only when the scope is understood. Then locate any conversation where the detail was originally typed or later repeated. OpenAI states that deleting a saved memory does not remove its mentions from past chats, while deleting the chat alone may leave a separately stored memory available.［1］

Use this two-part checklist:

Remove or correct the memory through the current memory controls.
Delete the source chat if its visible transcript should no longer remain in the account.
Check files or Library items separately if the chat included uploads; chat and file retention can differ.［4］
Record the action without copying the sensitive value into your log.
Follow the published retention policy rather than claiming instant backend erasure.［1］［4］

Deleting is not the same as archiving. An archived chat remains in the account and follows the retention rules for saved chats. Use archive for organization, not for a deletion request.

Step 4: Choose future controls for each type of work

Turning memory off can prevent ordinary future personalization, but it does not erase previous chats. OpenAI also warns that if memory is re-enabled, remaining chats may contribute memories again under the applicable experience.［1］ Choose the control based on the task:

Task	Safer default
Routine preferences	Memory on only for approved low-risk details
One-time sensitive discussion	Minimize or omit the data; consider Temporary Chat if policy allows
Long-running workstream	Approved Project with scoped files, instructions, members, and retention
Confidential organizational work	Organization-approved account and administrator policy
Secret or regulated record	Do not place it in the chat; use the authoritative protected system

Data Controls deserve a separate review. Disabling model improvement can change how eligible new chats are used, but it does not remove chat history or replace memory cleanup.［2］ A network privacy tool such as AethoVPN can protect traffic on an untrusted network, but it cannot decide what ChatGPT remembers, delete account data, or enforce workspace permissions.

Step 5: Use ChatGPT Temporary Chat with precise expectations

Choose Temporary Chat before entering task content, not after a normal chat has already recorded it. Before the conversation begins, confirm both that Temporary Chat is active and whether it is non-personalized. If existing memories, Custom Instructions, or plugins must remain outside the task, do not choose the personalized option. Keep inputs minimal and avoid GPT actions or connected services unless their separate data handling is acceptable.［3］

Temporary Chat reduces ordinary history and does not create new memories while it remains temporary, but personalized mode can reuse existing memories. It is not permission to submit material you are not authorized to share. OpenAI documents limited safety retention, and data sent through third-party actions can follow the recipient's policy.［2］［3］ Custom or workspace controls can also affect the experience. Do not save the chat unless you accept its conversion to a regular conversation.

If the result becomes valuable, move only a reviewed, sanitized conclusion into an approved record. Do not preserve the original sensitive prompt merely because the output was useful. The cross-platform chat organization guide offers a neutral handoff format for decisions and sources.

Step 6: Perform fresh-chat verification

After changing controls, start a new ordinary chat and run a fresh-chat verification with non-sensitive probes:

Ask what broad preferences are remembered without restating the deleted fact.
Ask for a recommendation in the affected category and check whether the old assumption returns.
Inspect the memory source indicator if the interface provides one.
Reopen settings and confirm the selected controls remain in the intended state.
Check that source chats or files targeted for deletion no longer appear in their respective account views.

A clean answer is evidence, not proof of complete backend deletion. Keep the account UI result, official retention statement, and any administrator confirmation separate. If the old detail appears again, look for another chat, Project instruction, custom instruction, uploaded file, or workspace source before repeatedly deleting the same memory.

Step 7: Establish a recurring review

Review memory after major life changes, role changes, Project sharing, account migration, or a policy update. A quarterly review is a reasonable starting cadence for active personal use; higher-risk work should follow the organization's schedule.

Use a minimal register with the category, decision, owner, review date, and evidence location. Never put the actual secret in the register. The ChatGPT safety overview can help assess adjacent risks such as links, files, shared chats, and account access.

Summary

Decide what is safe to reuse before changing a setting.
Treat memory, chat history, training controls, files, and Temporary Chat as separate layers.
Remove both a stored memory and its source conversation when both contain the unwanted detail.
Use task-local or governed project context instead of account-wide memory for changing work facts.
Verify in a fresh chat without reintroducing the information.
Follow official retention language and escalate secrets to their real owner for rotation.

Frequently asked questions

Does deleting a ChatGPT chat delete its memory?

Not necessarily. OpenAI documents that legacy saved memories are stored separately from chat history. Delete the relevant memory and the source chat when both should be removed.［1］

Does turning memory off erase old chats?

No. Turning off memory changes personalization behavior; it is not a chat-deletion action. Review chat history, files, and retention separately.［1］［4］

Should I ask ChatGPT to remember a password?

No. Passwords, recovery codes, API keys, session tokens, and similar secrets should remain in an approved secret manager. If one was exposed, rotate it rather than relying on chat cleanup.

Is Temporary Chat completely unrecorded?

No. An unsaved Temporary Chat stays out of ordinary history and is not used for model training, but OpenAI may retain a copy for up to 30 days for safety. A personalized Temporary Chat can use existing memories, Custom Instructions, and plugins, although it does not create new memories while it remains temporary. Saving it converts it into a regular chat.［2］［3］

Can I keep a harmless preference in memory?

Yes, if it is accurate, durable, low-risk, and useful across conversations. Add scope when a preference applies only to a project or a limited period.

Why did an old detail appear after I deleted a memory?

It may still exist in a chat, Project instruction, custom instruction, uploaded file, or another memory source. Inspect those layers without restating the sensitive value.

Does disabling model improvement delete my data?

No. That control governs whether eligible conversations help improve models; deletion and retention are separate controls.［2］

How often should I review ChatGPT memory?

Review it after important personal, role, workspace, or policy changes and on a recurring schedule. The right frequency depends on the sensitivity and rate of change of your work.

Disclaimer: Platform controls, availability, and retention terms can change and may differ for managed workspaces. Verify the current OpenAI documentation and your administrator's policy before handling confidential, regulated, or legally significant data.

Sources

OpenAI, “Memory FAQ” — https://help.openai.com/en/articles/8590148-memory-faq
OpenAI, “Data Controls FAQ” — https://help.openai.com/en/articles/7730893-data-controls-faq
OpenAI, “Temporary Chat FAQ” — https://help.openai.com/en/articles/8914046-temporary-chat-faq
OpenAI, “Chat and File Retention Policies in ChatGPT” — https://help.openai.com/en/articles/8983778-chat-and-file-retention-policies-in-chatgpt
OpenAI, “Projects in ChatGPT” — https://help.openai.com/en/articles/10169521-projects-in-chatgpt

Sources checked 4 October 2026.

How to Manage ChatGPT Memory and Clear What It Remembers