6 prompts that show what ChatGPT remembers about you, and how to delete it
OpenAI's own help page says the memory summary is not everything. These six prompts inventory what is stored, rank the risk and build a deletion list you can actually finish.
In this article
ChatGPT's memory is designed to disappear into the background. You mention a health issue once, a client's name, a child's school, and months later the answer arrives already knowing. That is the feature working. It is also how a private detail ends up living in a place you never look at.
OpenAI's help centre page on memory was updated this week, and it is unusually blunt about the gap. The memory summary is a "high-level view" and it does not include everything ChatGPT remembers. You can ask the assistant what it remembers, but even that answer is a summary of a summary. The only reliable way to see the shape of it is to pull the list out and read it line by line.
These six prompts are built for that job. Each one was run once against the same ten-item sample memory list before publication, and each produced a usable, structured answer. Keep the ones that fit your account and skip the rest.
What OpenAI's page actually says
Four details matter more than the rest.
First, memory can draw on more than saved memories: past chats, custom instructions, files in Library and content from connected apps such as Gmail, depending on your plan and region. Second, deleting a chat does not necessarily delete a separate saved memory created from that chat. Third, turning Memory off does not delete your past chats, and turning it back on can create new memories from the chats that remain. Fourth, removal has a tail: changes take time to propagate, and OpenAI may keep logs of deleted saved memories for up to 30 days for safety and debugging.
There is also a useful escape hatch. Temporary chats do not create or update memories, as long as you start them unpersonalized and do not save them afterwards. Saving a temporary chat turns it into a regular chat, and memory can use it from then on.
Before you start
Open Settings, then Personalization, then Memory. Depending on your account you will see some mix of Memory, Reference saved memories, Reference chat history, Memory summary, Manage and Saved memories. Open the summary and copy it into a plain text file first. That copy is your working list, and it is also the thing you will paste into the prompts below.
Two cautions. The summary may omit details, so a clean audit of it is not proof that nothing else is stored. And the prompts are written to be pasted into the same ChatGPT account you are auditing, not into a random third-party tool.
1. The inventory
What it does: turns a loose list into categories and a sensitivity rating, so you can see the pattern instead of ten separate surprises.
Here is my ChatGPT memory summary. Sort every item into identity, work, health, money, relationships, preferences. Rate each High, Medium or Low sensitivity, then tell me the three to delete first. My summary: [paste]
Tested: on a ten-item sample it produced a clean table, rated a child's school and a medical detail as High, and named three deletion candidates with a reason for each. The ratings are judgement, not policy, so override them freely.
2. The leak test
What it does: forces the model to argue against you. Instead of "is this sensitive", it asks what a stranger could do with it.
Act as a privacy reviewer. Read my memory summary and list the five details that would damage me most if a stranger read them. For each, name the exact risk in one line. Summary: [paste]
Tested: it returned five ranked risks, including a child's school, a medication, and a client name paired with a monthly fee. The useful part is the pairing. A name alone is boring; a name plus a price is a scam script.
3. The delete map
What it does: stops you from deleting one copy and assuming the job is done.
For every sensitive memory I list, show each place it can live: memory summary, saved memory, past chat, Library file, connected app. Then give the exact settings path to remove it. Sensitive items: [paste]
Tested: it produced a table across all five locations and the matching settings paths, and it was honest that it cannot inspect your account, so it marked the other copies as possible rather than confirmed. Treat that honesty as part of the answer.
4. The redaction rewrite
What it does: keeps the usefulness of a memory while removing the identifying detail, so you can replace rather than only delete.
Rewrite each memory below as a generic preference that keeps the same usefulness but drops names, places, employers, health details and money figures. Keep each under 12 words. Memories: [paste]
Tested: "client X pays Y per month" became "prefers tracking recurring client payments", and a diagnosis became "prefers personalised wellness guidance without retaining medical specifics". Slightly bloodless, but that is the point: the assistant still helps, the file stops identifying you.
5. The standing rule
What it does: writes the instruction that stops the next batch from being saved in the first place.
Write a 70-word custom instruction I can paste into ChatGPT that stops you saving my health, money, employer, location and family details, while still remembering my writing style and language preferences.
Tested: it returned an instruction at exactly 70 words that names the categories to refuse and the preferences to keep. It also added the caveat that matters: a custom instruction is not a deletion request, and it is not a guarantee. Memory behaviour still depends on your settings and plan.
6. The monthly report
What it does: gives you a repeatable review instead of a one-off panic.
Turn my memory summary into a one-page report: item count per category, the top three risks, and a keep or delete column. Flag anything that looks outdated. Summary: [paste]
Tested: it counted items by category, flagged a medication and a school as the top risks, and marked an undated workplace conflict as ambiguous rather than fact. That last habit is worth keeping: stale context is its own kind of wrong answer.
The deletion order that actually works
Work from the smallest container to the largest. Delete the individual saved memory or summary item first, then the chat where you first shared the detail, then any Library file that contains it, then disconnect the connected app if you no longer want it reading that mailbox or calendar.
Only after that consider the blunt controls. "Don't mention this again" changes behaviour but does not delete the source. "Delete and turn off memory" clears what the summary shows and switches Memory off, but leaves your chats. Turning off Reference chat history schedules remembered past-chat information for deletion within 30 days, while the original chats stay where they are.
The limits
This is housekeeping, not a guarantee. The summary is incomplete by design. Deleted items can linger in logs for up to 30 days. Workspace accounts may have administrator controls that override your personal settings, and Healthcare or regulated workspaces handle memory differently. If you are on a personal account with "Improve the model for everyone" turned on, chats and remembered information may be used to improve models; that switch lives in Data controls, not in Memory.
The realistic goal is not an empty file. It is a file you have read, pruned, and set rules for, with a monthly prompt that tells you when it has quietly grown back.
Our articles help you explore AI and build practical skills. For a correction or a question, contact us. Contact
