● A free public tool  ·  Built by Poptani Labs  ·  No sign-up  ·  Your data never leaves your browser
AI, Decoded · 5-minute read

Why your chatbot now remembers you — AI memory, explained

Screenshot of a chatbot comparison interface showing two AI responses side by side
Image: 'Chatbot Arena main UI.png' by LMSYS Org (UI), PrincessPandaWiki (prompter), "enigma" and Phi-4 (LLMs), Public domain, via Wikimedia Commons

Not long ago, every chatbot conversation started from zero. You could spend a week telling an assistant about your job, your writing style, and your kids' names — and the next chat would greet you like a stranger. That has changed. ChatGPT and Google Gemini now carry things across conversations: preferences, projects, and facts you mentioned weeks ago resurface unprompted. It feels like the assistant finally knows you. It also raises a question worth asking before you enjoy it too much: what exactly is it remembering, where does that memory live, and how do you take it back?

This is a research-based explainer, drawn from the providers' own documentation and independent privacy guides — not from hands-on testing of every setting.

What 'memory' actually means

Chatbot memory is not one thing. OpenAI describes ChatGPT's memory as an optional personalization feature that retains useful context across conversations: things like how you like answers formatted, a project you are working on, or a preference you stated. You can review, edit, or delete saved memories, and you can switch memory off entirely. Beyond explicitly saved facts, OpenAI has also expanded personalization so eligible versions of ChatGPT can reference information from previous conversations.

Google takes a different route with Gemini. There, memory is tied closely to Gemini Apps Activity: your chats and related activity can be kept for 3, 18, or 36 months, or retained indefinitely, depending on your setting. Temporary chats in Gemini are retained for 72 hours and are not used to train models. The details differ, but the shape is the same: the assistant is no longer a stranger each time, because your past conversations have become part of its working context.

Sources: What happens when AI remembers everything about you?, AI privacy settings and memory, compared

Memory, training, and history are three different switches

The single most useful thing to understand is that the privacy settings are not one switch but several, and they do not move together. History is whether you can find a conversation again. Memory is whether facts from it carry into new chats. Training is whether the provider may use your content to improve its models. Retention is how long a copy survives on a server after it disappears from your screen.

That separation creates surprises. Turning on memory does not turn on training, and disabling training does not stop an assistant from saving memories about you — you have to check both. Deleting a chat does not necessarily delete what memory learned from it. In ChatGPT, saved memories are stored separately from chat history and remain until you delete them individually in the memory settings; in Claude and Gemini, deleting chats also removes the associated memories. If you want a clean slate in ChatGPT, you need to clear both places.

Private modes have their own fine print. ChatGPT's Temporary Chats do not create new memories or appear in chat history and are not used for model improvement, though a copy may be retained for up to 30 days for safety. Gemini's temporary chats are kept for 72 hours. And retention promises have limits: a US legal dispute in 2025 temporarily forced OpenAI to retain even deleted chats, a reminder that a company's policy is not a law of physics.

Sources: What AI secretly remembers about you, AI privacy settings and memory, compared

The anecdote that says it all

We have been confiding in chatbots since long before they could remember anything. In 1966, MIT scientist Joseph Weizenbaum built ELIZA, a program of about two hundred lines that merely reflected users' words back as questions. His own secretary — who had watched him build it and knew exactly what it was — asked him to leave the room so she could talk to it in private. Weizenbaum later wrote that extremely short exposures to a simple program could induce powerful delusional thinking in quite normal people. Sixty years later, the programs actually do remember. The impulse to confide arrived first; the memory came second.

Source: IEEE Spectrum — why people demanded privacy to confide in the first chatbot

What to check in your settings

If you use a chatbot regularly, a five-minute audit is worth it. First, open the memory or personalization settings and read what is actually saved — delete anything stale, wrong, or too personal. Second, check the training control separately from memory: in ChatGPT it sits under Data Controls, in Gemini it is tied to keeping your Apps Activity. Third, use temporary or private chats for anything sensitive, and remember their retention windows are days or weeks, not zero. Fourth, keep genuine secrets out entirely — passwords, card numbers, and other people's confidences do not belong in a chat window, remembered or not. And if you share an account or device, remember that memory is personal only as long as the account is.

The same logic applies to the assistant's other habits: a chatbot that invents facts with total confidence, as covered in why AI sometimes makes things up, is a bad place to store anything you cannot afford to lose — and a password manager, not a chatbot, is where secrets belong, as the evidence on password managers shows.

What This Means

AI memory is genuinely useful — an assistant that remembers your preferences saves you from re-explaining yourself every day. But it is not a human memory: it is a database entry on someone else's server, governed by settings that move independently and policies that can change. Treat it like a notebook you lend to a company, not a confidant. Read what it has written about you, delete what you would not want repeated, and keep the truly private things out of the chat entirely. The assistant knowing you is a feature. You deciding what it knows is the skill.

More from AI, Decoded

← All posts