Does AI remember what you told it? How AI memory really works

Does AI remember what you told it?

No. AI has no memory at all. At every message, the tool sends the AI the whole conversation again, from the first word, and the AI re-reads all of it.

You send a first message, the AI answers. You send a second message: the AI receives your first message, its first answer, then your second message. At your fiftieth message, the AI does not see one message: the AI re-reads your 49 previous messages, its 49 answers, every document you attached and every web page it opened, then your new request.

Picture a consultant with total amnesia, to whom an assistant hands the complete file before each question. The consultant re-reads the whole file every time: the thicker the file, the more your question gets lost in it.

Why does AI get worse in a long conversation?

Because your new request becomes a needle in a haystack. The AI re-reads everything, and your latest ten words weigh almost nothing against the rest.

A rule you gave at the start ("address the client as Mr. Dupont", "never more than 100 words") sits, after an hour of work, under tens of thousands of words. Following one rule in a short paragraph is easy; following the same rule buried in a long history is much harder. That is when the AI starts forgetting rules, drifting from the topic and inventing details (see the guide Why does AI make things up?).

The same goes for a short reply at the end of a long exchange. The AI asks "Does this plan work for you, or should I change the order?" and you answer "yes": that single word arrives at the end of a history of several thousand words, and the AI has to guess what your "yes" answers. Reply with a full sentence: "The plan works, keep this order."

How does AI remember you from one conversation to the next?

Through notes that the tool adds to every new conversation. The AI itself still remembers nothing.

Many tools keep an account memory: when you mention something the tool judges important (your job, the breed of your tortoise), the tool writes it into a note about you. Your custom instructions ("answer briefly", "call me by my first name") work the same way. Both are plain text, added at the top of every new conversation, before your first word.

So you never start from zero. Before your first word, the AI has already received the tool's own instructions, your account memory and your custom instructions, and the AI re-reads all of it at every message, like the rest of the conversation. A memory note that grows over the months makes every conversation a little heavier.

Some tools can also search your past conversations when you refer to them ("remember what I told you about my tortoise?"). That search is an extra ability, like searching the internet: without it, each conversation stands on its own.

You can read and delete what the account memory says about you, in the tool's settings. Read that note once: you may find old or wrong information in it, which the AI re-reads in every conversation.

Can you ask AI to forget something?

No. Write "forget what I said": the AI will answer that it has forgotten, and everything will stay in the conversation, re-read at every message.

Worse, your "forget" message joins the history too: the AI now re-reads the information and the request to forget it. To make the AI forget, open a new conversation. To delete something from the account memory, go through the settings.

This matters most for what you should not have pasted: a client's name, a phone number, a contract. Once sent, the information stays in the conversation. The good habit comes before: send only what the task needs, and write [client] instead of the name (see Can I paste client data into AI?).

When should you give a document to the AI?

At the step that needs it. A document attached to the first message is re-read at every message; the same document attached at the tenth message did not weigh on the nine before.

Picture sending a whole dictionary to ask for three definitions: the AI re-reads thousands of definitions at every message to use three. Sending one entry at a time gives the AI only what each question needs. The same goes for ten reports, a long contract or a whole website: give each part when its step comes.

How do you start fresh without losing your work?

Tidy the desk, never the cabinet. The conversation is the desk: clear it as often as you can. Your files are the cabinet: they stay.

Deleting a conversation deletes neither the files on your computer nor the documents you produced with the AI and saved. So the rule is simple: one task, one conversation. When the task is done, open a new conversation for the next one.

When you need to carry on the same work, do not drag the whole history along. Ask for a summary, and say what the summary must contain:

Open a new conversation, paste the summary, and continue. The summary becomes the new starting point: a few hundred words instead of the whole history. Some tools offer a button that writes this summary for you; choose the moment yourself, because a summary made in the middle of a delicate step loses exactly the details you still need.

Why does a long conversation cost more?

Because AI use is measured by the amount of text the AI reads and writes, and a long conversation makes the AI re-read more at every message.

AI does not count words: AI counts tokens, pieces of words. A short word is often one token; a long word is often cut into several. Every token the AI reads, and every token the AI writes, has a cost: in money on a pay-as-you-go plan, in usage limits on a subscription.

Three facts follow, and they explain most limits reached too fast:

• Each message costs more than the one before. At every message, the whole history is read again: the fiftieth message costs far more than the first, even if both are ten words long.
• Writing costs more than reading. Text the AI writes costs several times more than text the AI reads. So a longer, precise request that gets a short, right answer costs less than a vague request followed by three corrections (see the guide How do you write a good AI prompt?).
• You pay for what you do not see. Before answering, many AIs reason: the AI writes a draft of its thinking, which the tool hides. That hidden draft is billed as writing, and it is often longer than the answer itself.

Two habits cost nothing. Paste text as text: a screenshot of a paragraph costs more than the same paragraph pasted, because the AI must first decipher the image. And English is the cheapest language for most AIs: the same request in French or Japanese uses more tokens.

Why does a conversation that sleeps cost more?

Because the tool keeps the work already done on your conversation for a short while only. Come back after lunch, or the next morning, and the whole history is processed again at full price.

While you keep working, the AI does not redo all its calculations on the history at each message: the tool keeps them aside for a short time (how long depends on the tool), and re-reading the history then costs much less. After a long pause, that work is gone. Your next message makes the AI process the entire history from the start, at full price, and the answer takes longer.

So a conversation that sleeps costs more than a conversation that moves. Finish the task before the break. If you cannot, ask for the summary before you leave, while the conversation is still warm, and start a fresh conversation with that summary when you come back.

What five habits keep the AI's memory under control?

Five habits to tick before a long work session with AI.

• One task, one conversation: when the task changes, I open a new conversation.
• I give each document at the step that needs it, not everything at the start.
• To make the AI forget something, I open a new conversation; I never write "forget".
• Before a break, I finish the task or ask for a summary, and I say what the summary must contain.
• I have read what my account memory says about me, and deleted what is wrong.

These habits keep the conversation light; they replace neither a good request nor a check. The two other guides cover requests and checks: How do you write a good AI prompt? and Why does AI make things up?