ChatGPT's memory summary misses details. 5 prompts to audit the list
A copy-paste privacy check-up: find what ChatGPT saved, fix the wrong lines, delete the rest, and keep a clean brief for temporary chat.
In this article
ChatGPT's memory is not one file. OpenAI's help centre says memory can draw on past chats, saved memories, custom instructions, files in Library and connected apps such as Gmail. The screen called the memory summary is only a high-level view, and OpenAI says it may omit details that are less relevant or come from sources it does not display individually. So the first move is not deletion. It is asking what the assistant thinks it knows, where each line came from, and what the summary left out.
That matters because the controls are split up. A saved memory and the chat that produced it are stored separately. A quiet button that stops the assistant mentioning something is not a delete button. And memory is not the same setting as model training. Getting those three distinctions right is most of the job.
Where the controls are
Open Settings, then Personalization, then Memory. Depending on your plan, region, platform and workspace, you may see Memory, Reference saved memories, Reference chat history, Memory summary, Manage or Saved memories. If a control described here is missing, OpenAI says it may not be available for your account yet. Workspace owners can also limit these settings for managed accounts.
Three distinctions that matter
First, "Don't mention this again" reduces future references to a detail. It does not delete the original chat, file, email or memory. Second, deleting a chat does not necessarily delete a separate saved memory created from it, and deleting a saved memory does not remove the chat where you shared the information. You may also need to remove it from Library files or a connected app. Third, memory controls are separate from training. The training switch is under Settings, then Data controls, and is called Improve the model for everyone. Turning it off stops new conversations being used for training, but it does not delete or hide saved chats.
The five prompts
We ran each prompt once in a clean chat and kept the five that returned a usable, specific answer. Paste them one at a time. Start with the audit, because you cannot correct a list you have not seen.
**1. Audit the list.** This asks for a table of memories, the model's confidence and a source, and forces an honest "unclear" when it cannot trace one. Treat the confidence column as the model's own estimate, not a fact.
List everything you remember about me in a table: what you remember, how confident you are, and where it came from. If you cannot name the source, write unclear. Mark anything I never confirmed.
**2. Correct the wrong lines.** This turns a correction into a replacement memory and asks for the exact place to paste it, so you are not leaving the old line in place beside the new one.
Here is what you got wrong about me: [old detail] should be [correct detail]. Rewrite it as one short memory in my voice, then tell me where to paste the correction and what it replaces.
**3. Prune both copies.** This sorts the list and, for each removal, names the chat, file or app you must also clear. That follows OpenAI's own warning that the saved memory and its source are stored separately.
Sort my memories into keep, fix and remove. For each remove, give me the settings path to delete it and the chat, file or connected app I must also delete, because the saved memory and the source are stored separately.
**4. Quiet what you want kept but not used.** For details that are true but sensitive, this writes the "don't mention again" line and states plainly what it does not delete, so you are not mistaking silence for removal.
These details are true but I do not want them used again: [detail]. Write one 'don't mention again' line for each, then list what that changes and what it does not delete.
**5. Keep a clean brief for temporary chat.** A temporary chat can be started unpersonalized, which means it does not use your memories or custom instructions. You choose personalized or unpersonalized before the conversation starts and cannot change it afterwards. This prompt gives you a short brief so the task still works without your profile.
Turn this into a 120-word brief I can paste into an unpersonalized temporary chat: [task, deadline, audience, constraints]. Leave out my name, contact details, employer and anything about money or health.
What we cut
Prompts that asked the model to "find every hidden data point" were cut, because the model cannot see storage beyond the sources OpenAI lists. Prompts that asked it to invent a source were cut for the same reason. The five above ask only for what the product exposes: the memory list, the sources panel and the settings paths.
The limits
The memory summary may leave out details, and a Sources panel may not show every factor behind a response. Deleted saved-memory logs may be retained for up to 30 days for safety and debugging. If you turn off Reference chat history, information remembered from past chats is scheduled for deletion within 30 days, but the original chats stay in your history unless you delete them. Temporary chats do not create memories and are not used to train models, but they may be kept for up to 30 days for safety. Self-service data export is available to signed-in Free, Go, Plus and Pro users, not to Business, Enterprise or Healthcare workspaces. A memory toggle also does not disable limited safety uses of context in rare, high-risk situations.
What to do now
Open the memory settings and ask for the audit. Correct the lines that are wrong, then delete the ones you do not want, then clear the chats, files and apps behind them. If you only want the detail ignored rather than deleted, use the quiet line and remember it is not a delete. Turn off model training in Data controls if you have not already. For the next sensitive task, start an unpersonalized temporary chat and paste the brief.
Our articles help you explore AI and build practical skills. For a correction or a question, contact us. Contact
