What your AI remembers about you (and what to stop telling it)

It’s 11pm. A letter has sat on the kitchen table all week — small print, a deadline. You’re tired, so you do the sensible modern thing: photograph it, paste in the numbers, ask your AI what to do.

It answers beautifully. It also, quite possibly, remembers.

Not the letter word-for-word. The shape of your life inside it: what you earn, what you owe, the health thing you’re worried about, your child’s school. Many AI tools now carry details like these from chat to chat, building a quiet little dossier that makes future answers faster and more personal — whether or not you meant to hand it over.

No reason to panic, or to stop using AI. Just give it the diary treatment: know what’s in it, cross out what shouldn’t be, decide what never gets written down.

First, three different things people call “memory”

Half the confusion around AI privacy comes from one word doing three jobs.

The chat itself. Every conversation sits in your history. Deleting it removes it from view — but not necessarily from memory.

Saved memories. Most major tools now also keep a short list of facts about you: your job, your kids’ ages, your preference for short answers. It’s what makes a new chat open with “How’s the weekend rush going?” instead of “How can I help?” Feels like magic; really a notebook.

Training. Separately, some of what you type in ordinary chats may improve the models themselves, under different settings. Turning memory off doesn’t change it. Three different drawers; tidying one leaves the others untouched.

The surprise: deleting a chat does not delete what the chat taught it. The notebook has to be cleaned out in its own place.

The 60-second audit

Before changing anything, just look. Ask your AI:

“What do you remember about me? List the personal facts you’ve saved, and anything you’re unsure about.”

Then prune. Most people find three kinds of entry: the useful (“prefers bullet points”), the outdated (“job-hunting” — that was eight months ago), and the “wait, I told you that?” The third kind is why you’re here.

Where the notebook lives, as of this writing:

• ChatGPT: Settings → Personalization → Memory — view, delete, or clear saved memories. Deleting a conversation does not remove memories it created; clear them here.
• Gemini: Settings & help → Personal context shows facts you asked it to keep (older labels may say Saved info). Chat history lives under Gemini Apps Activity / Keep Activity, and past-chat personalisation is controlled separately.
• Claude: Settings → Memory lists memories by topic; each can be viewed, edited, or deleted. Memory availability has depended on plan, so it may not appear for everyone.
• Microsoft Copilot: Look under Personalization (in some clients, Profile → Privacy → Personalization) for saved memories; or simply ask “What do you know about me?” and say “Forget that…”. Availability differs between personal and work or school accounts.

If the chat answer sounds vague, trust the settings list — chat-history-derived context may not enumerate like a saved-memory list.

Honesty note: these features are young, roll out gradually, and the labels move. If a menu name doesn’t match, search your app’s settings for “memory” — you’ll land in the right neighbourhood.

Three things that never get pasted

You don’t need a privacy policy for your kitchen table. Three bright lines will do.

1. Identity and financial documents. Passports and ID cards, bank statements, tax returns, payslips, account numbers. If a document would let someone pretend to be you or reach your money, the AI does not need the real thing. The safe workaround: redact first, then ask. Cover names, numbers, and balances, and describe the situation in your own words — “a loan at this rate, over this many years” earns the same explanation with none of the exposure.

2. Your children’s records and details. School reports with full names, medical letters, anything pairing your child’s identity with their health, school, or struggles. Kids can’t consent to a dossier, and shouldn’t inherit one built before they could read. The safe workaround: strip the identity, keep the question. “A 9-year-old finding fractions hard” is all the AI needs to build a practice game. It never needed the school’s letterhead.

3. Other people’s private information. A colleague’s performance review, a customer’s complaint with their phone number, a relative’s diagnosis told to you in confidence. It isn’t yours to hand over — and once it’s in your account’s memory, it’s in yours. The safe workaround: summarise, don’t forward. Retype the situation with names swapped for roles: “a team member,” “a long-time customer.” AI works just as well on the pattern as on the person.

All three workarounds share a secret: the AI almost never needs the document. It needs the situation. Describe the situation.

Whose memory is it, anyway?

Memory attaches to an account, and households share accounts like a fridge. One tablet, one login, means one person’s memories answering everyone’s questions — your teenager’s homework chat can surface your confidences, and theirs can surface in yours.

Five minutes fixes most of it:

• One account per adult, at minimum. Children use an adult’s account with that adult, or not at all.
• Agree the three bright lines out loud. Unspoken rules are just hopes.
• Prune seasonally. Clearing chat history? Clear the memory list too — different drawers.
• Work accounts are not private accounts. Personal matters belong on your own account; client matters don’t belong on either without permission.

A shared account builds a shared memory — and nobody in the house voted for that.

Reality check

“Temporary” and “incognito” modes are genuinely useful — and routinely misunderstood.

ChatGPT’s Temporary Chat, Gemini’s Temporary Chat, and Claude’s Incognito chats make roughly the same core promise: the conversation stays out of your visible history, doesn’t create new memories, and isn’t used to personalise future chats or train models. Depending on the tool and the mode, though, a temporary chat may still draw on existing memories or profile preferences you already have. For a one-off question you’d rather not have following you around, that’s still the right tool.

One more caveat: Claude’s Incognito chats aren’t saved to history or memory, but Claude can still access profile information such as custom styles and personal preferences inside an incognito chat — so it isn’t a completely blank identity.

What they do not promise is instant invisibility. These chats are still processed on company servers and retained briefly afterwards — OpenAI says up to 30 days, Google up to 72 hours — and providers reserve the right to review for abuse or safety, and to keep material where the law requires. Temporary mode is a raincoat, not a vanishing act: it stops the conversation sticking to your account, not from ever existing. Never type anything into a temporary chat you’d be devastated to see attached to your name — and keep redacting either way.

And one unfashionable truth: the strongest privacy setting is the sentence you don’t send. Every toggle in this article manages information already shared. The bright lines matter more.

The notebook is yours to edit

Used well, memory is lovely — no more re-explaining your life to a fresh chat every morning. The goal was never an AI that knows nothing about you. It’s one that knows what you chose to tell it.

So tonight, run the audit. Cross out what surprises you. Keep the notebook — and hold the pen.

Treat your AI’s memory like a diary you lend out: decide what gets written in it, and check the pages now and then.