A model has no memory in the way you do. It doesn't carry yesterday around with it, and it doesn't keep a running file on you. Each time it answers, it works from one block of text, assembled fresh and handed over all at once. Everything the model can consider lives inside that block. The block is called the context window.
What goes in it? More than people expect. Your instructions go in. Your question goes in. So does every earlier message in the conversation, any document you pasted, any results a tool fetched, and, as it writes, the answer so far. The window is measured in tokens (the tokens note covers those), and every model has a maximum. Current windows are large enough to hold a long book. They're still finite, and the cost of using them is real, because you pay for everything in the window each time it's sent.
Here's the analogy. Imagine a meeting room with one whiteboard and a rule: the person facing the board can only act on what's written there. Nobody may bring notes, and nobody can phone a colleague. At the start of the meeting, you write the brief up. As the discussion goes on, the board fills. Once it's full, the only way to add something is to wipe something else. After that, a question about whatever you wiped gets a puzzled look, because as far as the person at the board is concerned, it was never said.
Notice what that doesn't describe: forgetting. The model doesn't forget the way you forget a name. The wiped material is gone from its world, with no fuzzy trace, no feeling of "I used to know this." That has three practical consequences.
First, long conversations drift. When a chat gets long enough that early messages are dropped or shortened by the software around the model, the model will cheerfully contradict something you settled an hour ago, because for it, nothing was settled.
Second, a fresh chat starts blank. Whatever you told it in one conversation, it brings nowhere else, unless the product you use copies notes into the new window for you.
Third, even inside the limit, more isn't automatically better. Models tend to use the start and end of a long input more reliably than the middle, and a window stuffed with loosely related text gives the model more places to wander. A short, well-chosen window usually beats a long, careless one, and costs less.
This happened to us
Our enrichment step reads a job posting and fills in structured fields: location rules, seniority, and so on. It runs on a small model, and a small model has a smaller window. Our instructions take up part of it. The posting has to fit in whatever remains.
Most postings do. Some don't: pages of company history, benefits boilerplate, legal notices, all wrapped around the few paragraphs that decide the answer. The obvious fix is to cut the text off wherever the window runs out. That keeps whatever happened to come first, which is a different thing from keeping whatever matters most, and the model would never tell us it had been handed half a posting. A truncated input looks complete from the inside.
So we trim by priority instead. When a posting is too long, the parts that carry the most information are kept first, and the filler is what falls away. The window is a budget, and we decide who gets spent from it before the model sees anything.
See it yourself (2 minutes)
Open a brand-new chat in your agent or any AI tool and send:
Then open a second, separate new chat and ask:
It can't know, and that second chat is a fresh whiteboard. Now go back to the first chat and ask it:
Compare its answer with what you've actually sent. You're looking at the whole of what the model knows about you right now.
What this means when you build
Every agent you build in this program has a window to budget: your instructions, the tools' output, the documents you feed it, the history of its own steps. A few missions in, you'll hit the moment where something important quietly stops being in front of the model, and the symptom will look like a model that got dumber. Before blaming the model, check what was on the board. You'll also learn to decide, deliberately, what gets written there first.