Yes, ChatGPT keeps a profile of you, and you can read most of it tonight. The catch is that the profile is a summary, not a full inventory of everything the system may remember.
This article gives you nine copy-paste prompts and a settings checklist. Each one turns a vague privacy worry into a specific list of what is stored, where it came from, and which control reduces the risk.
You do not need to code. You need a browser, about 30 minutes, and the willingness to open your own settings instead of trusting a guess about what is in them.
Before You Start
1. Sort Your Stored Memories Into One List
Prompt to copy
Act as my privacy auditor. Using the memory summary I am about to paste, group every remembered detail into these categories: work, money, health, family, preferences, locations, and sensitive identifiers. For each item, mark whether I explicitly told you or whether it looks inferred. Flag anything I would not want used in ads or model training. End with the five details I should correct or delete first. Here is my memory summary: [PASTE MEMORY SUMMARY].
On the web, click your profile icon, then Personalization, then Memory, and choose Manage next to Memory summaries. The phone app hides the same page under the menu button and your profile icon. Copy the whole text into the prompt box.
About 40 seconds after you paste, the reply sorts every line into seven buckets, flags each as told or inferred, and names five items to fix first.
That list is a starting point. OpenAI calls the summary high-level, not a full inventory, so budget 15 minutes to check it against your own chats.
π‘ Tip: OpenAI says the memory summary is a high-level summary, not a complete inventory. Treat it as a starting map, not a guarantee.
2. The Wrong-Memory Correction Prompt
Correcting the saved summary does not rewrite the chats behind it. Old conversations stay, and they can be re-indexed if memory is switched back on.
So fix the lines first, and do not expect a clean slate from the edit. A grouped summary shows the categories, not which lines are false.
The old way means scrolling every line and guessing which ones still apply. The new way pastes the summary into this prompt and lets the assistant mark each one.
Prompt to copy
Review the memory summary below as a fact-checker. Mark each line as confirmed, stale, ambiguous, or wrong. Rewrite stale and ambiguous lines into precise replacements I can paste back into memory. Do not invent details. Ask me one question at a time when you need confirmation. Finish with a short list of lines I should delete rather than correct. Memory summary: [PASTE MEMORY SUMMARY].
The answer takes about 30 seconds and returns four labels plus rewrites you can paste. Review eats the time here, roughly 15 minutes. Then open "Manage", highlight a line, and choose "Make a correction" or "Don't mention this again" to save each fix.
β οΈ COMMON MISTAKE
Editing the summary is not deletion
Changing a line in the memory summary edits the saved profile, not the underlying chats. If you need the detail gone from the conversation too, delete that chat separately.
3. The Inferred-Details Audit Prompt
Allow about 40 seconds for the answer and 10 minutes to judge it. A line can be true and still feel like an intrusion.
So the question worth asking is not only whether a detail is right, but whether the reader ever said it out loud.
Inferred habits sit behind stated facts, and they are the ones that make a recommendation feel too close.
Prompt to copy
Act as an inference auditor. Read the memory summary below and list only the details that look inferred rather than directly stated: routines, mood, health, finances, relationships, location, religion, politics, or shopping habits. For each, say what clue in the summary may have led to it, how confident you are, and whether it is sensitive enough that I should delete it. Do not add new facts. End with the three inferences most likely to make an ad or recommendation feel intrusive. Summary: [PASTE MEMORY SUMMARY].
The catch is that ChatGPT cannot reliably describe its own inference process. The list may be incomplete, and parts of it may be invented to fill the gaps. The check is to compare each inferred item against the actual summary and your chats.
Read the output as a set of hypotheses, not a confession. It stays on patterns in the summary you pasted, not hidden system data.
You still get items tagged with a possible clue, a confidence level and a sensitivity call, plus the three inferences most likely to make an ad feel intrusive.
π‘ Tip: Turn on temporary chat before running this one: use the dotted speech bubble at top right on the web. It commits nothing to memory and is forgotten when closed.
4. The Shared-Links Exposure Prompt
Sorting a link list by eye takes about 20 minutes across two years of chats. Pasting the titles and dates into ChatGPT returned a rated list in about 15 seconds. The review, checking each rating against the actual chat, took 9 minutes.
Open Settings, then Data controls, then Shared links. Every conversation you have ever shared sits there with a title and a date. Copy those two fields only. The full chat text never goes into the prompt.
The study behind the worry: in May 2026, IMDEA Networks researchers tested nine browser assistants and eight Android apps. Every service contacted at least one advertising or analytics company.
They counted 44 companies on the receiving end. Shared conversation links ranked as the most common leak.
Prompt to copy
Act as a link-risk reviewer. I will paste the titles and dates of my shared ChatGPT links. For each one, rate the exposure as low, medium, or high based on whether the title suggests health, money, work, family, legal, or identity details. Tell me which links to delete first, what a recipient could infer from the title alone, and what I should never put in a shared link. Do not ask for the full chat text. My shared links: [PASTE TITLES AND DATES].
Back comes a three-level rating per link, a top-to-bottom deletion order, and a guess at what the title alone gives away. A title like "MRI results explained" reads high even if the body is dull.
Deleting a link stops new visits. It does not pull back a copy or a screenshot someone already saved. By hand: about 20 minutes. With ChatGPT: about 15 seconds, plus 9 minutes to check.
π‘ Tip: Deleting a link blocks new visits, but it does not remove copies or screenshots others already made. Treat deletion as risk reduction, not recall.
5. The Uploaded-File Memory Prompt
Finding what an uploaded file left behind means checking two places. The title of a chat may say "budget.xlsx" while the memory summary quietly holds a line about your contract terms.
Tracing that by hand means scrolling months of chat history one thread at a time.
References to files can survive in the memory summary and chat titles even when the file contents are not listed separately. So paste only those, never the files themselves.
Uploaded contracts, medical notes and financial records can hold client data, account numbers or diagnoses. Never paste passwords, client files, or medical or financial records into a chatbot.
Prompt to copy
Act as a file-retention reviewer. From the memory summary and chat titles I paste below, identify any references to uploaded files, including PDFs, spreadsheets, images, contracts, medical records, or financial documents. For each reference, say what type of file it appears to be, what personal data it may contain, whether it belongs in memory, and what I should delete or keep out of future chats. Do not invent file contents. End with a three-line rule for what I will never upload again. My summary and chat titles: [PASTE].
The review names each file type, the data it may carry, whether it belongs in memory, and what to remove or keep out of future chats.
It closes with three rules for what never gets uploaded again. For anything you cannot bear to store, answer in a chat you delete instead.
π‘ PRO TIP
Title-only audit is enough
For the shared-links and file prompts, paste only titles and dates. The reviewer can rate exposure without the full text, and you avoid creating a second copy of the sensitive material.
6. The Model-Training Opt-Out Prompt
On free and personal plans they may be, unless you switch that off in settings. So this control decides where your words travel, not just what sits in memory.
A Canadian joint investigation found OpenAI's initial ChatGPT training was not compliant with privacy laws. OpenAI later limited personal and sensitive information used to train new models. Read that as context, not proof that current settings are safe.
Because setting names and locations change, the prompt writes a checklist with placeholders rather than guessing a menu path. Verify each step in the product's own Settings and Data controls before you trust it.
Prompt to copy
Act as a settings-checklist writer. Create a plain-English checklist for a US ChatGPT user who wants to stop their chats from being used to improve the company's models. Include: what to look for in Settings and Data controls, how to tell whether the toggle is on or off, what to do if the wording is confusing, and how to record a screenshot and date for my own records. Do not claim to know the exact current menu names; use placeholders like [SETTING NAME] and tell me to check the product's own help pages. End with a five-item checklist I can complete in under ten minutes.
Back comes a five-item checklist covering what to look for in Settings and Data controls, how to read the toggle, what to do with confusing wording, and how to date a screenshot. The AI draft takes seconds. Running it takes about 8 minutes.
7. The Ad-Profile Shutdown Prompt
Model-training settings govern one thing: whether new models learn from your chats. Ads are a separate track. OpenAI's US privacy policy says it collects ad history and interests from Free and Go users, such as the types of ads they view or interact with.
Niedersachsen's state data protection commissioner Denis Lehmkemper warns that using earlier conversations as well as the current one for personalized ads can build very detailed personality profiles. Consent lives in your account's current privacy settings, not in a chat command.
Paste the ad section of the privacy policy under a prompt that reads it back in plain words.
Prompt to copy
I want to understand what my account's privacy settings allow regarding ads. Below is the ad section of the privacy policy [paste the ad section of the privacy policy here]. Read it back to me in plain words. Produce four things: (1) the ad and interest data it lists, (2) what consent changes according to the pasted text, (3) a short script for checking my settings, and (4) a place for me to record my choice. Keep it to about a minute of reading and stay strictly within the pasted text.
Expect four things: the ad and interest data listed, what consent changes per the pasted text, a short script for checking settings, and a place to record your choice. It takes about a minute to produce.
Your check: confirm each item against the policy text you pasted.
No universal toggle exists, and paid plans and regions differ, so the setting may be worded differently or sit elsewhere on your account.
If you cannot find it, record that too, with a screenshot and date. Settings check: consent, ad-free options and the date you checked them.
π‘ Tip: Screenshot the toggle state and the date. If the wording changes later, you have a record of what you chose.
What Each Control Actually Does
π§ Memory toggle off
Stops new memories
ποΈ Memory toggle off
Keeps old memories
π¬ Temporary chat
Saves nothing
π Delete shared link
Blocks new visits
π₯οΈ Computer History
Off by default
The pattern
Most controls stop future collection. Few of them erase what is already stored.
8. The Temporary-Chat Test Prompt
Two minutes buys a habit that never stores a question: temporary chat. On the web, start one with the dotted speech bubble at the top right. Mobile entry may differ, so the app's own screen is your check.
Nothing in it goes to memory, and it is forgotten once closed. It is not a cloak: the privacy policy still covers the chat. A short setup makes the boundary clear before you need it.
Prompt to copy
Act as a temporary-chat guide. Explain in plain English what temporary chat does and does not save, when it is the right mode for a sensitive question, and when I should skip chat altogether and use another channel. Then give me three test questions I can ask in temporary chat to confirm it behaves as expected, and a one-line note about when to close and start a new temporary chat. Do not claim to know my account's exact settings.
What a Control Stops vs. What It Leaves
What it stops
- Turning off memory stops new memories from being saved
- Deleting a shared link blocks new visitors
- Temporary chat saves nothing to memory
- Screenshotting a setting records your choice
What it leaves
- Existing memories and chat history remain
- Copies and screenshots others made remain
- The conversation is still governed by the privacy policy
- Computer History memory files stay local until removed
9. The Delete-and-Start-Clean Prompt
Turning off memory stops new memories. It does not wipe the ones already stored, and it does not delete a single past chat.
Cleanup has to run as an ordered pass across memories, chat history, shared links, uploads, ad preferences, training settings and any local Computer History files.
Deletion and export sit in the product's own Data controls. Use those, never a guessed menu path. Account type and country change what applies, so the planner prompt asks for both and leaves setting names as placeholders to verify yourself.
Prompt to copy
Act as a data-deletion planner. Build a step-by-step cleanup plan for my ChatGPT account using these categories: saved memories, chat history, shared links, uploaded files, ad preferences, model-training settings, and local Computer History files if I use the macOS desktop app. For each category, tell me what deleting it does, what it does not do, the order to do it in, and what I should screenshot as proof. Add a final check for anything that would remain after I turn memory off. Use [PRODUCT SETTING] placeholders where the exact name may change and tell me to verify in the product's own Data controls. My account type is [FREE / PLUS / PRO / BUSINESS / ENTERPRISE], and my country is [COUNTRY].
The plan comes back in seconds. Expect, per category, what deleting does and does not do, the order, and the screenshot to keep. Fix any step that assumes a menu name before you run it.
Computer History, where it applies, is an opt-in macOS feature for Pro, Business and Enterprise. It is off by default and unavailable in the EEA, Switzerland and the UK.
It keeps temporary event files up to 48 hours and leaves memory files local in plain text until you remove them.
The final check matters most: ChatGPT's memory FAQ says safety features may still use limited context in rare high-risk situations. Screenshot each step with its date.
The result is a dated record of what stopped and what remains, which is the honest boundary of a cleanup.
The Privacy Terms You Will See
Frequently Asked Questions
Can I see everything ChatGPT remembers about me?
You can see the memory summary, but OpenAI says it is a high-level summary, not a complete inventory of everything ChatGPT may remember. Use it as a starting map and check the underlying chats for details that do not appear in the summary.
If I turn off memory, are my old memories deleted?
No. Turning off memory stops new memories from being saved, but existing memories and chat history stay. Past conversations can also be re-indexed if memory is turned back on.
Do temporary chats save anything?
Temporary chat commits nothing to memory and is forgotten when you close it. The product's privacy policy still governs the conversation, and the entry point can differ between the web app and the mobile app.
Is ChatGPT using my chats for ads?
On US Free and Go plans, OpenAI's privacy policy says it collects ad history and interests, such as the types of ads you view or interact with. Check your current privacy settings for personalized-advertising consent and ad-free options.
What is Computer History and should I worry about it?
It is an opt-in feature for the ChatGPT desktop app on macOS, for Pro, Business and Enterprise, and it is off by default. It records clicks, typing and app switches, keeps temporary event files for up to 48 hours, and leaves memory files local in plain text.
Does deleting a shared link remove the conversation?
No. Deleting a link blocks new visits, but copies or screenshots others already made remain. Delete the underlying chat as well if you need the conversation gone from your account.
Run the Inventory First, Delete Last
Run the Memory Summary Inventory prompt first. Then correct the wrong lines, audit your shared links and uploads, and check the model-training and ad settings. Finish with the Delete-and-Start-Clean prompt, so deletions come last.
Hold on to two limits. The memory summary is a high-level summary, not a complete inventory. And deleting a chat is separate from turning memory off.
Worth doing anyway: you end with named settings, prompts you can rerun, and a dated record. Every control you touch limits future storage or access. None of them is a recall button.
Start where you started reading: the profile is a summary, and now you have read it, corrected it and dated it.