The goldfish problem: why every chat starts from zero
Close the tab and an AI forgets your child completely - name, class, the maths problem from ten minutes ago, all of it. Here is why that happens and what it teaches a child about the tool.
A ten-year-old told us her name, her class and her favourite subject at the start of a session. Four minutes later, in a fresh chat window she had opened by accident, she asked the same AI what her favourite subject was. It had no idea. She checked the screen twice, convinced something was broken.
Nothing was broken. That is simply what closing a chat and opening a new one does. Every conversation is its own sealed room, and nothing said in one room is heard in the next.
Why every chat starts from zero
A large language model does not carry a running diary of everyone it has ever spoken to. Each conversation is, technically, a fresh calculation that only sees what is written inside that same conversation window. Close the tab, and the words inside it are gone from the model's point of view - not deleted from a server somewhere necessarily, but no longer visible to the next conversation you start.
We call this the goldfish problem in class, because it is the closest thing to how children already picture a goldfish's memory - three seconds, then a clean slate. It is a slightly unfair metaphor for a goldfish, but it lands instantly with a ten-year-old.
The test that proves it in under a minute
- Tell the AI your name in a chat.
- Close that tab completely.
- Open a brand-new conversation.
- Ask: "what is my name?"
Every single time, the answer is some version of "I don't know - you haven't told me." Children find this more convincing than any explanation, because they ran the experiment themselves.
Why this is not actually a design flaw
It is tempting to hear "no memory" and think the tool is unfinished. It is closer to a safety choice than a limitation. A model that silently remembered everything typed into it, across every device and every family member, would be a far more worrying product than one that forgets on purpose between sessions.
Forgetting on purpose is not the same as being broken. Sometimes it is the feature working correctly.
What actually solves the goldfish problem
Re-typing your name, class and preferences at the top of every chat works, but it is tedious, and tedious things get skipped. Level 3 of this pathway solves it properly: write the About Me facts down once, in a place that gets attached to every new conversation automatically, so the goldfish problem stops being your child's problem to manage by hand.
Questions we get asked
Why does the AI forget what I told it?
Each conversation with most AI chatbots is a separate, sealed session by default. Nothing typed in one conversation is visible in a new one you start afterwards, which is why an AI can seem to instantly forget a name or fact you gave it minutes earlier in a different chat.
Is AI's lack of memory a bug that will get fixed?
It is largely intentional rather than a fault. A model that silently remembered every family member's conversations across every session, with no visible control over what it kept, would raise far bigger privacy concerns than one that starts fresh each time. Some products now offer memory as an optional, visible setting instead.
How do you stop having to repeat the same information every chat?
Write it down once as a short block of context - name, class, how you like answers explained, any rules - and attach it to the start of every new conversation. This is exactly what level 3 of our pathway, context engineering, teaches a child to build.