{"id":1492,"date":"2026-09-01T22:02:20","date_gmt":"2026-09-02T01:02:20","guid":{"rendered":"https:\/\/prompteiro.com\/en\/what-chatgpt-keeps\/"},"modified":"2026-09-01T22:06:54","modified_gmt":"2026-09-02T01:06:54","slug":"what-chatgpt-keeps","status":"publish","type":"post","link":"https:\/\/prompteiro.com\/en\/what-chatgpt-keeps\/","title":{"rendered":"What ChatGPT Keeps of Everything You Type"},"content":{"rendered":"

You paste a whole contract in to understand one clause, photograph a blood test to find out what the acronym means, drop your CV in there to sharpen the description of a job. In none of those moments do we stop to think about where any of it went. The doubt arrives later: is this being kept somewhere?<\/p>\n

It is, and the full answer is duller and calmer than the fear suggests. In this lesson you will see the three different things people call “keeping”, how to empty the memory, what temporary chat is actually for, and the short list of what is not worth pasting in there. About 8 minutes of reading.<\/p>\n

The golden rule: once you type it, it has left your device<\/h2>\n

Keep this sentence, because it organises everything else: what you send to an AI has gone to another company’s server<\/strong>, exactly the way an email attachment goes. There is no drawer with your name on it inside the model. There is a service that received, processed and logged what passed through it. The lesson on what artificial intelligence is<\/a> covers that “inside” calmly.<\/p>\n

Out of that comes a test that beats reading any privacy policy. Before pasting, ask: would I email this to a supplier I have never met in person? If yes, paste without drama, and that covers most things. If no, the problem is not the tool, it is that piece of the text, and the fix is usually deleting two lines.<\/p>\n

Three different things we all call “keeping”<\/h2>\n

Almost all the confusion here comes from mixing up mechanisms that are independent: they have separate switches, and turning one off does not turn off the other.<\/p>\n

\n\n\n\n\n\n
The mechanism<\/th>\nWhat it does<\/th>\nWhat you control<\/th>\n<\/tr>\n
History<\/td>\nkeeps the conversation in the side list so you can come back to it<\/td>\ndelete one conversation at a time, or the whole account<\/td>\n<\/tr>\n
Memory<\/td>\nwrites notes about you and uses them in future conversations<\/td>\nsee, edit and delete each note, or switch it off entirely<\/td>\n<\/tr>\n
Training<\/td>\nuses what went through to improve the next versions<\/td>\na switch in the settings, which you can turn off<\/td>\n<\/tr>\n
Review<\/td>\nallows a person to read an excerpt where abuse is suspected<\/td>\nnothing, and that is why the rule above is the rule<\/td>\n<\/tr>\n<\/table><\/div>\n

The first two rows are yours and you can sort them out in two minutes. The third is a switch almost nobody has opened, and it tends to come on in personal accounts and off in company ones, which already answers half the questions about what changes between paying and not paying<\/a>. The fourth is the one nobody controls, and it is what turns the golden rule into a real rule.<\/p>\n

Memory and temporary chat: the two switches that fix this<\/h2>\n

Memory is the newer feature that confuses people most and the one most worth knowing. Instead of forgetting everything when the conversation ends, the assistant writes notes about you and uses them later: that you have two children, that you work in accounting, that you prefer short answers. It is genuinely useful, which is why plenty of people only discover it exists when it fills up and the warning appears.<\/p>\n

Do three things today. Open the list of memories and read it<\/strong>, which is the part that both alarms you and solves the problem. Delete the notes that should not be there, and you will find at least one. Then decide whether to keep memory switched on, knowing the tool works well without it.<\/p>\n

One thing almost everybody gets wrong: deleting the history does not empty the memory, and emptying the memory does not delete the history. They are different switches on different screens, and touching one gives you the feeling of having sorted out the other.<\/p>\n

Temporary chat is the second switch. It does not enter the history and does not feed the memory, so it works for the question you would rather not see in the side list: the health worry, the conversation about your boss, the idea that is not anyone’s yet.<\/p>\n

What it is not: a mode in which nothing leaves your device. The message reaches the server exactly the same way and is usually held for a period for abuse checks. Temporary chat is the browser’s private window: it deals with the trail that stays with you<\/em>, and it does not make you invisible on the other side.<\/p>\n

Photos and files say more than you typed<\/h2>\n

With text you choose word by word. With photos and files you do not: they go whole, with everything that was in the frame and everything the file carries inside it. The gap got wider once the camera became the fastest way to ask for something, which is the subject of the lesson on using AI on your phone<\/a>.<\/p>\n

The commonest case is the photo of an identity document. You only want to improve the portrait, and you use Turn a Passport Photo into a Proper Portrait<\/a> for that, but the photo that came off your phone has the whole document in the frame: number, parents’ names, date of birth. Cropping before you send takes three seconds, and the result improves, because the AI starts looking only at the face.<\/p>\n

\"Turn<\/a>
Turn a Passport Photo into a Proper Portrait \u2192<\/a><\/figcaption><\/figure>\n

Files hide the same trap: the spreadsheet goes with the tabs you forgot about, the document goes with the comments and the edit history. When what you need is the content and not the file, pasting the passage works better. A Faithful Summary of a Long Document<\/a> handles pasted text just as well, and then only what you read first goes out.<\/p>\n

What you can paste without thinking twice<\/h2>\n

This deserves saying plainly, because the subject attracts exaggeration: most of what we do with AI has no personal data in it at all, and freezing up about everything is the easiest way to never use the tool.<\/p>\n

Public text is the clearest example. The leaflet you photographed is printed in every box of that batch, so running it through Understand Your Medicine Leaflet in Plain Language<\/a> exposes nothing of yours. What does not go with it is the prescription, which has your name and your doctor’s on it.<\/p>\n

A standard clause is the same story. The Legalese Translator<\/a> works on the loose paragraph, with no header and no parties, and the result does not get any worse: that text is identical in thousands of contracts.<\/p>\n

What deserves a black marker first<\/h2>\n

There is a whole category of useful task involving documents with numbers on them, and not one of them needs the number to work. Deleting the header takes ten seconds.<\/p>\n

Understand Your Payslip<\/a> wants the deduction lines, not your employee number. Understand Your Tax Refund<\/a> wants the amounts and the dates, not your tax ID. In both cases, paste from the middle down.<\/p>\n

Health needs more care, because the lab header carries full name, date of birth and patient number on the same line. Make Sense of Your Lab Results<\/a> only needs the values and the units. And the usual reminder applies: this is reading, not a diagnosis.<\/p>\n

With contracts the marker improves the result as well as protecting you. Both the Contract Analyser<\/a> and Rental Contract: Check Before You Sign<\/a> do better with “landlord” and “tenant” in place of the real names: the analysis becomes about the clause and stops getting distracted by the parties.<\/p>\n

And then there is what we forget because it is routine. The Ruthless CV Reviewer<\/a> is one of the most used prompts on the site, and every CV starts with name, phone number and address. The criticism that matters is about what you delivered, and it comes out identical without those first four lines.<\/p>\n

\"Ruthless<\/a>
Ruthless CV Reviewer \u2192<\/a><\/figcaption><\/figure>\n

Your face is data too<\/h2>\n

A photo of your face is the most personal data most people send without noticing they are sending it, because a photo does not look like a form. And the rule here is not “never”, it is which photo<\/em>.<\/p>\n

When the image is headed somewhere public, the maths is easy. Profile Picture for WhatsApp and Social<\/a> works on the photo you were going to publish anyway: it leaves your device to become exactly the thing you wanted the world to see.<\/p>\n

The maths changes when it is a photo you would never publish: the one taken inside your home, the one with an address on the sign behind you, the one from a private moment. It is not forbidden, it is a conscious decision, and it gets easier when the edit works on a crop of just the part that matters.<\/p>\n

The data that is not yours<\/h2>\n

This is the section almost nobody writes, and the only one where the consequence lands on somebody else. About your own data, you decide how much you want to risk. About someone else’s, you are deciding for a person who was never asked.<\/p>\n

At work this comes up daily. Meeting Summary<\/a> is excellent, and the notes you are about to paste have client names, proposal values and information your employer considers its own. The right question here is not whether it is safe: it is whether your company has already said what is allowed.<\/p>\n

Photos of other people follow the same logic, and the most delicate example is children. Restore a Baby and Childhood Photo with AI<\/a> is one of the loveliest uses there is when the photo is of your own family and the decision is yours. When it is a pupil, a nephew or a patient, the person with rights over it decides.<\/p>\n

What is never worth sending<\/h2>\n

The list is short, and short on purpose: passwords, two-step verification codes, full card numbers, system access keys. It is not about trusting the company. It is that this data exists precisely so it does not leave where it lives, and one of them pasted into a conversation stays in the history.<\/p>\n

That does not stop AI from helping with the subject. A Strong Password You Can Remember<\/a> was written with exactly that lock built in: it teaches the passphrase method and the order in which to protect accounts, and it says outright that you are not to type any real password there. A good security prompt is one that does not ask for what it does not need.<\/p>\n

One last thing: answers about privacy age fast, because settings move around. Check on your own device whatever the AI tells you about where each switch lives, using the routine from the lesson on why AI gets things wrong<\/a>.<\/p>\n

The five minute routine to do today<\/h2>\n

This settles almost everything the text discussed, in order of effort:<\/p>\n

    \n
  1. Open the list of memories and read it.<\/strong> Delete what should not be there and decide whether to keep memory on. Biggest relief for the least work.<\/li>\n
  2. Find the training switch in the settings<\/strong> and choose deliberately. The point is not which side, it is having chosen.<\/li>\n
  3. Delete the old conversations you will never reread.<\/strong> A clean history is less exposed material if somebody gets into your account.<\/li>\n
  4. Turn on two-step verification for your AI account.<\/strong> It now holds as much about you as your email does, and that is where the real risk lives. A Strong Password You Can Remember<\/a> builds the plan.<\/li>\n
  5. Make cropping a habit.<\/strong> Crop the frame before sending a photo of a document, and delete the header before pasting any paper with a number on it.<\/li>\n<\/ol>\n
    \"A<\/a>
    A Strong Password You Can Remember \u2192<\/a><\/figcaption><\/figure>\n

    Quick questions<\/h2>\n

    Does ChatGPT keep my conversations?<\/h3>\n

    It does, by default, and they sit in your account until you delete them. Deleting a conversation takes it off the list, and copies may remain in the company’s systems for a period, which is how any service with a recycle bin is built.<\/p>\n

    Does what I write become training for the model?<\/h3>\n

    It can, and there is a switch for it in the settings. It tends to be on in personal accounts and off in company ones. It is worth opening and choosing rather than leaving it on the default.<\/p>\n

    My ChatGPT memory is full. Now what?<\/h3>\n

    Filling up is normal, because the note space is limited. Open the list, delete what no longer makes sense and the space comes back. Use the clear-out to read what is in there: it is the fastest way to find out what the tool has noted about you.<\/p>\n

    Is temporary chat safe?<\/h3>\n

    It deals with the trail on your side: it does not enter the history and does not feed the memory. It is not anonymity, because the message reaches the server all the same and is usually held for a while. It is for the conversation you would rather not keep, not for data that must not leave.<\/p>\n

    Can I send a document with my ID number or bank details?<\/h3>\n

    You can, but you almost never need to. None of the common tasks needs the number to work, so delete the header and paste the rest. Passwords, verification codes and full card numbers stay out every time.<\/p>\n

    Does deleting the conversation delete what the AI learned about me?<\/h3>\n

    No. History and memory are separate things on separate screens. Deleting the conversation does not touch the memory notes, and clearing the memory does not delete the history. To be properly clean, do both.<\/p>\n

    Where to start today<\/h2>\n

    Open the list of memories before you close this page. Two minutes, you will read things you had forgotten telling it, and you come away with the right feeling: it is not that the tool is dangerous, it is that it is a service, and services are things we configure.<\/p>\n

    After that, the text prompts hub<\/a> has the prompts this post cited, the image prompts hub<\/a> has the ones that start with a photo of yours, and the course series<\/a> carries on from here. Privacy is not deciding between using and not using: it is deciding what goes in, and that gets decided once and becomes a habit.<\/p>\n","protected":false},"excerpt":{"rendered":"

    Does ChatGPT keep your conversations? The difference between history, memory and training, and what is worth covering up before you paste.<\/p>\n","protected":false},"author":1,"featured_media":1491,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_i18n_grupo":1951,"footnotes":""},"categories":[64],"tags":[],"class_list":["post-1492","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-courses"],"_links":{"self":[{"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/posts\/1492","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/comments?post=1492"}],"version-history":[{"count":1,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/posts\/1492\/revisions"}],"predecessor-version":[{"id":1493,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/posts\/1492\/revisions\/1493"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/media\/1491"}],"wp:attachment":[{"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/media?parent=1492"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/categories?post=1492"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/tags?post=1492"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}