The Goldfish Problem: Why AI Forgets (and Why That's Your Problem Now)

The Goldfish Problem: Why AI Forgets (and Why That's Your Problem Now)

Picture this. It's a Tuesday evening. You've opened a chat window with your favorite AI assistant and asked it to help you plan a two-week trip through Japan in October. The first message you send is something like: *"I'm thinking about going to Japan for two weeks in October. I love historical sites, I don't love huge crowds, my budget is moderate, and I'd like to mix cities with some quieter places. Can you help me sketch an itinerary?"*

The AI fires back a beautiful response. It picks shoulder-season dates to avoid the worst crowds. It suggests Kyoto, Kanazawa, Takayama, and a few days in a rural onsen town. It explains why each stop matches your tastes. It even flags that mid-October is typhoon tail-end season and offers a Plan B. You think: this thing is brilliant.

You reply: *"Great, can you swap one of the cities for somewhere with great food?"* The AI suggests replacing one stop with Osaka, then seamlessly rebuilds the itinerary to fit. Still brilliant.

You ask: *"What about transportation between these cities? I'm not renting a car."* The AI walks you through the Japan Rail Pass, regional lines, and a couple of bus options. Still brilliant.

Turn four: *"Are there any ryokans you'd recommend in the onsen town you mentioned?"* The AI lists three, each with a one-liner about why it suits you. Brilliant.

Turn five: *"What's the weather usually like in Kyoto in mid-October?"* The AI gives you a layered answer: average highs and lows, rainfall, what to pack. Brilliant.

Turn six. You write: *"Okay, can you put the whole updated itinerary into a clean day-by-day schedule I can show my travel partner?"*

And the AI replies: *"Sure! Where are you thinking of going, and when?"*

You stare at the screen. You just spent thirty minutes building something together. You scroll up. The conversation is right there. Every preference, every swap, every clarification, all visible above. But the AI is asking you where you want to go, as if it has never met you.

This is the goldfish problem.

It happens to almost everyone who uses AI seriously. It happens to executives who paste in a long email thread and ask for a summary. It happens to developers who feed an AI a thousand lines of code and ask why it keeps hallucinating functions that don't exist. It happens to doctors, lawyers, students, customer support agents, and, increasingly, to the people building these systems. The AI seems terrifyingly capable in one moment and completely clueless in the next. The shift is so abrupt that it feels personal, as if the AI has decided to stop paying attention.

It hasn't decided anything. The AI doesn't have a choice. The problem is mechanical, structural, and a little bit embarrassing for the entire field: most of the time, when an AI seems to have "forgotten," it has done so because of a hard physical limit in how it processes your conversation. That limit is called the context window, and understanding it is the single most useful thing you can learn about modern AI. Because once you see the workbench, you can never unsee it. And once you can see the workbench, ninety percent of AI weirdness stops being mysterious and starts being predictable.

This chapter is the opening move of a book about that workbench. We are going to spend the rest of our time together figuring out how to lay things out on it carefully, because the people who do that well — who understand what fits, what doesn't, what to keep, what to throw away, what to summarize, what to retrieve — those people are quietly becoming some of the most valuable professionals in any industry that touches software. The job doesn't even have a settled name yet. Some call it *prompt engineering*. Some call it *context engineering*. Some call it *AI product design*. The term we will use most often in this book is the one that, in 2025 and 2026, has begun to win the argument: context engineering.

But before any of that, we need to slow down and look carefully at what context actually is. We need to define our terms. We need a mental model strong enough to carry us through the next fourteen chapters. And we need it without a single intimidating diagram.

Almost.

The Workbench Is Fixed, and That Changes Everything

Let's build the workbench.

Imagine a sturdy wooden workbench in a busy workshop. On it, a craftsperson is going to assemble a complicated clock. They have all the parts laid out: gears, springs, the face, the hands, the case, a printed instruction manual, a notebook with notes from previous attempts. The workbench has to hold all of this, at once, while the craftsperson works. There is a limit to how much fits. If the parts spill onto the floor, the craftsperson cannot reach them, and so for all practical purposes, those parts do not exist.

Now imagine the workbench is exactly four feet wide. That is the only constraint. It doesn't matter how brilliant the craftsperson is. It doesn't matter how many years they have been building clocks. If the parts don't all fit on the four-foot surface, the clock will not be assembled correctly.

An AI's context window is that four-foot workbench. It is the fixed-size workspace in which the model reads everything it needs in order to produce its next response. Every word you have ever typed into that chat, plus the system's instructions, plus the AI's own previous replies, plus any documents you have pasted in, plus any tools the AI has been told about — all of it has to fit on the workbench at the same time, or it effectively does not exist for the AI. There is no magic shelf the AI can reach up to. There is no filing cabinet the AI can quietly slip into. There is only the workbench, and the workbench has a width, and the width is the width.

The first time you really feel this limit is the first time you paste in something long. Say you paste in a forty-page contract and ask the AI to summarize it. If the contract is too long, the AI does not get a "table of contents" and read just the headings. It does not skim. It does not remember the first page once the last page arrives. It either fits or it does not. If it does not fit, one of several things happens: the conversation may fail with a polite error message; the AI may silently truncate the start of your contract so that the *end* of the contract is what it sees; or, worst of all, the AI may produce a confident, fluent summary that mixes up details because it only ever saw half of what you gave it. That third outcome is the dangerous one, because it looks like intelligence. It is not. It is a four-foot workbench that only had room for half a clock.

Here is the part that surprises almost everyone the first time they hear it. The size of the workbench is not a moral choice the model makes. It is not the company being stingy. It is a mathematical consequence of how the underlying technology works. Under the hood, every modern large language model — and we will unpack what "large language model" really means in Chapter 3 — processes text in chunks called tokens, and it processes all of those tokens simultaneously through a mechanism called the transformer. The transformer's fundamental operation requires it to compare every token to every other token in the input. If you double the input, you quadruple the comparisons. If you quadruple the input, you multiply the comparisons by sixteen. The math gets expensive fast, and the cost is not just dollars; it is also the model's ability to *find* the right piece of information in a sea of other pieces. As the workbench fills up, the AI doesn't just get slower. It gets distracted.

This is the second surprise. Even before the workbench is full, the AI can start losing its grip. Modern research has consistently shown that as you stuff more information into a context window, the model's accuracy at any specific task tends to degrade. This is sometimes called "lost in the mid