What happens when you ask an AI to forget what you just told it (and why it can't)

There's a particular moment in many AI conversations where someone realises they've said something they'd rather take back. Maybe you've accidentally pasted in sensitive information, or you've gone down a rabbit hole that's now hopelessly muddled, or you've simply changed your mind about the entire direction of the chat. The natural human response is to say something like "actually, ignore that last bit" or "forget what I just told you about the database passwords."

Perfectly reasonable request. Completely ineffective result.

The thing is, asking an AI to forget something you've just told it is a bit like asking someone to unhear a joke. You can both agree to pretend it never happened, but the information has already done its work. The difference is that whilst a human might genuinely manage to put something out of their mind, an AI doesn't actually have a mind to put things out of. What it has instead is a context window, and that's an entirely different beast.

How context windows actually work

When you're chatting with an AI, everything you've said in that conversation exists in what's called the context window. Think of it as the AI's working memory for that specific chat session. Every message you send, every response it generates, every correction you make – all of it gets appended to this growing document that the AI references each time it formulates a response.

Here's the important bit: the AI doesn't have a concept of "remembering" or "forgetting" in the way humans do. It doesn't form memories or consciously recall information. Instead, every single time you send a message, the AI reads through the entire conversation from the beginning (or at least, as much of it as fits in its context window) and then generates its next response based on all of that text.

So when you say "ignore my previous message," what you're actually doing is adding another message to the context. The AI now has to process: your original message, plus your instruction to ignore it. It can certainly try to follow that instruction – most AI systems are trained to be helpful and accommodating – but the original information is still right there in the text it's reading.

Why 'forgetting' is really just 'pretending'

Modern AI language models are quite good at playing along when you ask them to ignore something. They'll do their best to act as though the information isn't there. But this is fundamentally different from the information actually not being there.

Imagine you're writing an essay and someone's reading over your shoulder, offering suggestions as you go. Halfway through, you write a paragraph that goes completely off-topic. You then write a note saying "please disregard the previous paragraph." Your helper will certainly try to pretend they didn't see it, but if your off-topic paragraph mentioned that you're actually writing this essay whilst sitting in a cafe in Prague, that context has already coloured their understanding of everything else you've written. They can't unknow it.

The AI is in the same position. It can attempt to act as though certain information doesn't exist, but that information still influences the statistical patterns it's using to generate responses. It's in the input, so it affects the output, even if the AI is trying very hard to be cooperative about your request.

What actually removes information

There's really only one reliable way to remove information from an AI conversation: start a new conversation. Clear the chat, open a fresh session, begin again. This gives you an actually empty context window rather than one that contains your information plus a polite note asking the AI to look the other way.

Most AI chat interfaces have a button for starting a new conversation, and if you've genuinely shared something you shouldn't have – actual passwords, personal information, proprietary data – that's the button you want to press. Don't rely on the AI to forget. It won't and it can't.

Some platforms also let you edit or delete previous messages in a conversation. This can be genuinely helpful, as it actually removes the information from the context window rather than just adding an instruction to ignore it. Though it's worth noting that this only affects your local conversation – if you've shared sensitive information, the better approach is to treat it as potentially compromised and change passwords or take other appropriate security measures.

The privacy implications

Understanding how context windows work matters for privacy and security. When you're using an AI service, you should assume that everything you type in a conversation is there for the duration of that conversation. Depending on the service and its privacy policy, it might also be stored on someone's servers, potentially used to improve the AI, or visible to moderators.

This doesn't mean you can't use AI tools for sensitive work – plenty of people do, and many AI providers offer enterprise versions with proper data protection guarantees. But it does mean you should be intentional about what you share. Don't paste in actual passwords, real personal data, or anything else you'd be uncomfortable having potentially read by a human moderator or incorporated into training data.

If you catch yourself about to ask an AI to forget something, that's usually a sign that you shouldn't have shared it in the first place. Start a new conversation instead, and this time keep the sensitive bits abstracted or anonymised.

Why this confuses people

The confusion is entirely understandable. We anthropomorphise AI because that's what humans do with anything that communicates with us. We use words like "remember" and "forget" because they're natural ways to describe how conversations flow. And AI systems are designed to sound conversational, which reinforces the illusion that we're dealing with something that thinks like we do.

But an AI doesn't remember your conversation the way a human colleague would. It doesn't have yesterday's chat lurking somewhere in its biological memory, growing fuzzy around the edges as time passes. It has a text file of your current conversation, which it reads from top to bottom every time you interact with it. Those are fundamentally different things, even if they can sometimes feel similar in practice.

The practical upshot

Next time you're in an AI conversation that's gone sideways, resist the urge to ask it to forget what you've said. Either start fresh with a new conversation, or just redirect the discussion without worrying about the previous messages. The AI will follow along either way.

And if you've accidentally shared something genuinely sensitive? New conversation, immediately. Then go change whatever password or information you've exposed. Don't trust the AI to forget, because it's literally incapable of doing so. The information is in the context, and it'll stay there until that context is cleared.

It's not a failing of the technology – it's just how it works. Once you understand that, you can use AI tools more effectively and more safely, without expecting them to behave like forgetful humans rather than very capable text-processing systems.

More on AI and automation

If you found this article useful, we've written quite a bit more about how AI actually works (as opposed to how people assume it works) and what you can realistically do with automation. These are good places to start:

• Understanding AI hallucinations and why they happen

• When to automate a process and when to leave it manual

• What AI agents can actually do in production environments

Why we write about this

We don't write about AI and automation as interested bystanders. This is the actual work we do. Modern software development increasingly involves AI – whether that's using AI agents to handle parts of the development process, building AI features into the applications we create, or helping clients understand what's actually possible versus what's just vendor hype.

We've built custom CRM systems that use AI to categorise and route customer enquiries, e-commerce platforms with AI-powered product recommendations that actually work, and internal tools that automate the sort of tedious data processing that used to consume someone's entire afternoon. The technology is genuinely useful when applied thoughtfully to actual problems.

You can read more articles right here:

Agentic Engineering Explained

Stop Overthinking Your Prompts

What 'hallucination' actually means when an AI confidently invents things that don't exist

Work with us

If you're thinking about building something – whether that's a website, a web application, or custom software that solves a specific problem your business has – we'd be happy to talk. We're particularly good at projects where AI or automation could be genuinely helpful, but we're equally comfortable telling you when it wouldn't be. To see some examples of what we build, take a look below.

TrailTrack - The Best Ever Outdoor Hiking Website

RiVault - Encrypted Password Manager

RiCreate - Content and Social Media Creator

We build things that work properly, we explain technical concepts without the usual nonsense, and we're allergic to promising features that can't actually be delivered. If that sounds like the sort of people you'd like to work with, get in touch and let's have a conversation about what you're trying to achieve.

Get In Touch

For prices, please see the top of the page!