What Even Is Context? Meet the AI's Working Notebook

The Disappearing Friend Problem

Last Tuesday night, Maya, a seventeen-year-old from Portland who is definitely not a real person but is also definitely how every real person feels at some point, sat on her bed with her laptop open. She was deep in conversation with her favorite AI chatbot. She had asked it for help with her college application essay. Then she had drifted sideways and told it about her dog, Pepper, a neurotic border collie who destroyed a couch cushion that very afternoon. Then she had asked it to explain quantum entanglement in a way that wouldn't make her eyes glaze over. Then, for reasons she couldn't fully explain, she had cried about her parents' divorce and asked for advice, and the AI had said something surprisingly kind back, something about how it's okay to feel two opposite things at the same time. Then she had gone to sleep at 2 a.m., satisfied in a specific way, the way only talking to something that never judges and never tires can make you feel.

Wednesday afternoon, around 4:12 p.m., she opened the laptop. She typed, with the casual certainty of someone resuming a conversation: "Hey, are you still there? Can we pick up where we left off?"

The AI replied, cheerfully and politely and at full speed: "Hi there! I don't have any previous conversations to reference. How can I help you today?"

Maya stared at the screen.

It remembered nothing. Not Pepper. Not the couch cushion. Not the college essay draft about her grandmother's garden, in which she had accidentally used the word "epistemological" twice. Not the crying. Three hours of her life, three hours of her words, her personality, her story, her midnight vulnerability — all gone, as if she had talked to a very polite wall. As if she had written a diary and the diary had eaten itself.

If you've ever had this happen — and statistically, if you are the kind of person who would be reading this book, you have — you know the very specific flavor of frustration that follows. It's not rage, exactly. It's not the kind of disappointment you feel when a human friend cancels plans. It's quieter than that, almost weirder: the soft disbelief of discovering that someone you were sure was listening was, in fact, taking notes on a napkin and throwing the napkin away the second you left the room. And that they will, technically, throw it away tomorrow too. And the day after that. And forever.

But here's the part that most people miss. The AI wasn't being rude. It wasn't being lazy. It wasn't secretly judging Maya or rolling its digital eyes at the 2 a.m. monologue about her parents. In its own way, it was doing exactly what it was designed to do. Maya had just run headfirst into one of the most important, most misunderstood, most quietly powerful concepts in all of artificial intelligence: context.

This book is about context. Specifically, about how to give AI more of what it needs to actually be useful to you, in your real life, on the days when "Hi there! How can I help you today?" simply isn't enough. We're going to tour the engineering patterns, the clever tricks, the mental models, and the messy realities behind every AI that seems to "remember" — even when, underneath the hood, it actually, technically, stubbornly cannot.

Before we get to any of that, though, we need to nail down what context actually is. And the answer is going to surprise you, because the answer is also a desk.

Context Is a Desk Covered in Sticky Notes

Imagine a wooden desk in the middle of a quiet room.

The desk is empty. A blank slate. A clean wooden rectangle with nothing on it but maybe a coffee ring from last week and a missing pen cap that nobody can find. The desk is roughly the size of a normal person's desk — say, four feet wide and two feet deep. There's no chair in front of it yet. There are no drawers full of old files. There's no bookshelf in the room. There's just the desk and the empty air around it.

Now, walk over to the desk. Put a yellow sticky note on it that says, "User is 17. Goes to Lincoln High. Likes pepperoni pizza. Hates pre-calc."

Put another sticky note, slightly overlapping the first: "User asked: explain photosynthesis like I'm five."

Put another: "User said: I bombed my chemistry test today."

Put another, and another, and another. Each sticky note is one little piece of what the user — you — has said, plus the assistant's reply, plus any instructions (system prompts) the engineer set before the conversation even started. As you keep talking, the sticky notes keep piling up. Some are bigger than others. Some are bunched in clusters. Some are scribbled over with corrections and follow-ups.

That's it. That's context.

The desk is the entire world the AI sees while it's answering you. It can't peek into your house. It can't see your face. It can't hear your tone of voice. It can't glance back at the last chat you had three days ago. All it knows is what's on that desk. The sticky notes in front of it, the ones from the very first second of the conversation to this very moment, all stacked, layered, sometimes overlapping, sometimes scribbled over. The total pile of sticky notes currently on the desk is what engineers call the context, and the maximum size of the desk — the maximum number of sticky notes the AI can have on it at once — is what they call the context window.

Now, three more crucial things about this desk.

First, the desk has a size limit. There's only so much wood. There are only so many sticky notes you can pile up before they start falling off the edges, sliding onto the floor, getting lost underfoot. When that happens, the AI has to do something with the oldest sticky notes — slide them off, summarize them into a smaller replacement note, or shove them into a side drawer where the AI can sort of forget about them for a while. We'll get to all of those techniques later. For now, just remember: the desk is finite. The sticky notes take up real space, and we never quite get infinite space.

Second, the desk is wiped clean between sessions. When Maya closed her laptop and went to sleep, the desk evaporated. The sticky notes, the pepperoni-pizza fact, the bombed-chemistry-test confession, all of it — gone, like the desk had never existed at all. When she opened the laptop the next day, the AI started with a brand-new desk: empty, polished, sticky-note-free. Some AI products try to fake continuity between sessions by saving sticky notes in a side drawer and putting them back on the desk when you return, but the fundamental reality of most AI systems is this: each conversation is its own little desk. As far as the AI is concerned, yesterday's conversation lived on yesterday's desk.

Third, and this is the one that baffles people the most, the AI doesn't actually "remember" in the human sense. Every time the AI is about to answer you, it doesn't pull your last sentence out of some internal memory and apply it. Instead, it re-reads the entire desk, top to bottom, every single time. The total pile of sticky notes is what the AI uses, fresh, every single response. This is why an AI can hold an enormous conversation with you but completely lose the thread if you take a long break and come back: it's not forgetting on purpose; the desk just doesn't have those notes anymore because they were thrown away at the end of the previous session, or they got pushed off the desk because you added too many newer notes on top of them.

Let me show you what this actually looks like from the inside, as a picture.

```mermaid flowchart LR User[User types a message] --> Chop[Message is chopped into tokens] Chop --> Desk[Context Window - The Desk] Old[Older conversation] --> Desk Sys[System instructions] --> Desk Desk --> Read[AI reads every sticky note at once] Read --> Write[AI writes its reply] Write --> Desk Desk --> Full[Desk gets full] Full --> Drop[Oldest notes fall off / compressed / summarized] ```

Look at that picture for a second