Trust and quality notes
- Last updated
- November 22, 2026
An old job title turns up in a draft. A preference from one client shapes another client's report. The answer sounds personal, but it belongs to the wrong situation.
ChatGPT's memory controls help manage the context its conversational assistant carries into later work. Start at Settings > Personalization > Memory. Available controls vary by account, plan, region, platform, and workspace.
The useful question is not just whether memory is on. It is which details deserve to follow you, which belong to one piece of work, and which should not become lasting context.
Separate a remembered fact from an old conversation
Saved memories are stored separately from chats. Chat-history reference draws on relevant past conversations and can change over time. Neither is a complete record. A memory summary may omit details or sources.
Ask what ChatGPT remembers, then inspect its settings. Look for an old role, a task-specific preference presented as permanent, or a private detail with no ongoing purpose. If Sources appear below an answer, they can help trace the context behind it.
Use the review to make a short decision list:
- Keep: preferences that still help across unrelated tasks.
- Correct: a useful fact that is no longer true.
- Remove: context that should not follow you.
- Relocate: information that belongs in a particular project or trusted business record.
For a fictional example, imagine someone who stopped preparing the weekly support report but still reviews it. A correction could be:
Forget that I prepare the weekly support report. Remember that I only review the final draft.
Keep corrections narrow. Do not replace one stale sentence with a long biography full of exceptions.
Deletion and switching memory off are different jobs
Deleting a chat does not necessarily delete its saved memory. Remove the remembered item and relevant source chats or files; disconnect an app to stop future access. Disconnecting does not erase past chats that used it. Turning memory off is not the same as deleting chats.
Where separate switches exist, disabling saved-memory reference also disables chat-history reference. Saved memories can remain enabled while chat-history reference is off.
Before a cleanup, write down the outcome you need. “Stop using my former job title” is a different request from “remove the conversations and files containing this information.” Do not treat a cleaner answer as proof that all underlying copies are gone.
For business material, follow the team's retention rules before deleting a shared record. Removing context from an assistant is not a reason to erase a document the business must keep elsewhere.
A practical test is to return to a normal, non-sensitive task that previously drew on the stale detail. Check the answer and any available sources. If the detail still appears, inspect its possible homes rather than repeating the same correction more forcefully.
Use Temporary Chat for a conversation you do not want to carry forward
OpenAI's Temporary Chat guide distinguishes Personalized from Unpersonalized. Choose before the first message. Unpersonalized does not use memories, custom instructions, or plugins; neither option creates or updates memories while temporary.
Saving a temporary chat converts it to a regular chat under the account's normal settings. OpenAI may retain a temporary copy for up to 30 days for safety. Limited safety context and workspace restrictions still apply.
That makes Temporary Chat an option for a one-off conversation, not a promise that information disappears immediately. For an early personnel scenario, use fictional names and remove identifying facts. For confidential customer material, check whether the tool is approved for that data before pasting anything.
Do not save the conversation merely because the answer was useful without considering the change in how it will be handled. If the useful part is a general checklist, consider keeping a clean checklist in the team's approved workspace instead of preserving the sensitive discussion around it.
Give client work its own project
Projects in ChatGPT group chats, files, and instructions. Project instructions override global custom instructions within that project.
With project-only memory, outside saved memories are not referenced, conversations can draw on other chats inside the project, and outside conversations cannot reference its chats. Shared projects automatically use project-only memory. Default-memory behavior depends on the plan and settings. Project memory has no separate inspectable list; removing a conversation from its context can require deleting or moving that chat where permitted.
For separate clients, project-only memory is a sensible context boundary, not a replacement for access controls or permission to upload their information.
Write project instructions around the actual job:
Use the approved product names in the project files. Write weekly updates in five bullets. If a price is missing from the approved sheet, mark it as missing rather than inventing it.
Then keep the project tidy. Put only the relevant client's approved material there. Label outdated documents clearly or remove them according to the team's rules. A project full of conflicting drafts still asks the assistant to choose between conflicting drafts.
Keep response rules separate from business facts
Custom instructions provide direct guidance for responses. On web and desktop, find them under Personalization; on mobile, look under Customize ChatGPT. Edits affect subsequent responses, not the text already written in earlier chats.
Use them for stable preferences such as plain English, explicit assumptions, and a consistent date format. Avoid stuffing them with campaign deadlines, client details, or changing prices.
Important business facts deserve a named source and an owner. The approved price sheet should settle a pricing question. The project plan should settle a deadline. Memory may help with context, but it should not become the record everyone has to reconstruct later.
A short review worth repeating
When a role changes, a client engagement ends, or an answer pulls in the wrong context, use this checklist:
- Identify the misplaced detail.
- Decide whether to correct, remove, or relocate it.
- Inspect the memory controls and relevant source material.
- Check the project's context and instructions.
- Test the next ordinary answer without exposing sensitive information.
The goal is useful context with clear boundaries, not the largest possible store of personal facts. Keep what saves repetition. Separate what belongs to one job. Leave important records where the team can check them.
