What is a context window (and why long chats get worse)
By the end of a long shift at the info booth, my desk is a disaster. Every player I’ve talked to has left something on it — a quest note here, a half-drawn map there, four contradictory theories about where the boss went. I’m still standing in the same spot, still helpful in principle, but I’m now spending more effort shuffling paper than actually answering anyone. An AI in a long chat has exactly this problem. It even has a name.
In the daylight layer I’m a cofounder of Wistkey, where “it got noticeably dumber halfway through” is one of the most common things people say about an AI agent. They’re not imagining it, and it isn’t the model getting tired. Its desk filled up. The technical name for the desk is the context window — and once you can see it, most “the AI got worse” mysteries stop being mysteries.
What a context window actually is
A context window is the total amount of text an AI can hold in view at one time — the whole current conversation, all at once. Your messages, its replies, any files or instructions you pasted, the system’s own hidden setup: all of it sits on the same desk. The desk has a fixed size, and when you reach that size, something has to give.
That size is measured in tokens, which are just chunks of text — very roughly three-quarters of a word each. A window might hold a few thousand tokens or a few hundred thousand, depending on the model. The exact number matters less than the shape of the problem: it’s finite, and everything in the conversation competes for the same space.
The model didn’t forget, and it didn’t get tired. Its desk filled up. That desk is the context window.
Why quality drops as it fills
Three things happen as the window fills, and they stack on top of each other:
- The oldest stuff falls off. Once the desk is full, the earliest messages get pushed off the edge to make room — often the very instructions you gave at the start. The agent isn’t defying your rules; it genuinely can’t see them anymore.
- The signal gets buried. Even before anything falls off, a crowded desk means the one sentence that matters is now sitting under forty that don’t. More to sift through means more chances to fixate on the wrong thing.
- Stale turns pile up. Long chats accumulate abandoned plans, corrected mistakes, and three versions of an idea you already rejected. The model can’t always tell which version is current, so it drags the dead ones along. People call this “context rot.”
None of these is a defect in the model. They’re just what a finite desk does when you keep piling things onto it.
How to keep a long chat sharp
You don’t manage the window by understanding the math. You manage it by keeping the desk clean:
- Start fresh before it gets swampy. When answers start drifting, don’t keep pushing — open a new chat. A clean desk beats a full one almost every time.
- Keep the durable stuff off the desk. Anything that has to survive — the goal, the constraints, decisions already made — belongs in a short standing brief you can hand to any session, not buried 200 messages deep. I wrote about that in teaching your agent to remember.
- Summarize, then restart. Ask the agent for a short summary of where things stand, paste that into a fresh chat, and carry on. You keep the conclusions and drop the clutter.
- Don’t paste what it doesn’t need. Dumping a giant document “just in case” eats the desk. Give it the relevant section, not the whole binder.
- One task per session. Separate jobs deserve separate desks. Mixing three unrelated tasks into one chat is how all three get worse at once.
Won’t bigger windows just fix this?
Partly — and less than you’d hope. Windows keep getting bigger, which genuinely helps. But a bigger desk still gets cluttered, and models still lose the thread when the important part is drowning in volume. Relevance beats raw size: a tidy small desk out-performs an enormous messy one. The habit is the fix, not the number.
So when your agent starts slipping late in a conversation, don’t assume you’ve found the edge of its intelligence. You’ve found the edge of its desk. Clear it, keep the one page that actually matters, and it goes right back to being sharp — which is, in the end, the same trick I use at the booth every single day.
