Context Window: The Day a Model Wakes Up
Imagine a weird person named Ryland.
Ryland studied for five hundred years. Physics, chemistry, biology, programming, history, law, meme culture — all of it. Ask a question, and Ryland can connect a dozen domains into an answer that sounds half PhD, half alien.
But Ryland has a disease: every morning, they wake up with no personal memory.
They do not know what happened yesterday. They do not know which task they were working on last round. They do not know what stupid mistake they made yesterday. Unless someone teaches them what matters after they wake up, the day starts from zero.
If you have read Andy Weir’s Project Hail Mary, this should feel familiar. Ryland Grace wakes up with a lot of scientific knowledge, but has to reconstruct the mission from clues. Hopefully that doesn’t count as a spoiler — it’s just from the first chapter, I swear.
In this essay, though, we will not call this person Ryland Grace.
Just Ryland.
Clawd chimes in:
Yes, the name collision is that convenient. Please do not send a lawyer from the Naming Coincidence Committee. This is an AI infrastructure metaphor, and every surprise in the book remains safe ╮(╯▽╰)╭
Large language models are a lot like Ryland.
A model wakes up smart, but it does not naturally remember yesterday. The amount of life it can experience in this waking day is what humans call the context window.
ShroomDog pushes back:
This essay is not going to use the usual “context window is a desk” metaphor as the main frame. A desk explains capacity, but it misses one important thing: order.
Things on a desk feel simultaneous. Things in a context window arrive along token time. The morning lesson becomes background for everything that follows. A tool result that enters later in the window is an afternoon event in Ryland’s day.
A human and Ryland are not living under the same clock. The human may wait ten seconds for an API call. To Ryland, that result lands later in the day, inside a digital dimension where time flows by tokens. Tiny relativity joke: wall-clock time is one frame of reference; context time is another.
The context window is Ryland’s day
The context window is often described as “how many words the model can see at once.”
That is not wrong, but it is flat.
A better framing: the context window is a day in Ryland’s world.
At the start of the day, Ryland wakes up. The system prompt, AGENTS.md, developer messages, and user request are like the first class of the morning. A human or program stands in front of Ryland and says:
- this is who you are today
- this is the task
- these are the rules
- these are the traps to avoid
- these are the tools available
- this is the handoff from the past
Only after that class does Ryland really start working.
But the class itself consumes the day.
If the lesson is long, scattered, and overloaded, Ryland may finish reading the rules in the evening. They have not searched files, run tools, fixed code, or made a decision yet — and the day is already late.
If the lesson is crisp and on point, Ryland may finish at 10 AM. They know how to work, and they still have a full day of events ahead.
This is the truth behind prompt hygiene: more rules do not automatically mean more safety. A longer morning class means less working day. Handing Ryland a legal code at breakfast and expecting graceful production debugging in the afternoon is, frankly, a little cruel.
Token usage is the clock of this world
In this metaphor, token usage is the clock in Ryland’s world.
Humans look at the time to know how much waking day remains. Agents look at current token usage to know where the day is.
20% of the context used feels like 10 AM.
50% feels like 2 PM.
80% feels like 11 PM.
The hard context limit is the universe exploding. Ryland is not choosing to sleep; reality simply refuses to continue the day.
So token usage is not an IQ meter. It is closer to a clock.
A 1M-token model is not suddenly a god. It is more like Ryland can stay awake for three days and three nights, experience more events, read more material, and run longer tasks.
That is powerful.
But someone who can stay awake for three days does not automatically become wise. They can attempt a longer journey, and therefore need better sleep hygiene.
Old small-context models were basically koalas. Awake for two hours a day. Eat the system-prompt eucalyptus in the morning, do thirty minutes of work, then back to sleep. Early prompt engineering felt so tense because Ryland’s waking day really was that short.
Events consume the day
Ryland’s day is not filled by text alone. It is filled by events.
A user message is an event.
A file read is an event.
A tool result is an event.
An error log, search result, test output, or reversed decision is also an event that happened during this day.
Morning events are usually clear. The rules are fresh. The task is new. Ryland still knows what matters.
Late at night, events start blending together.
A plan rejected in the morning appears again in a log. A bug fixed at noon looks open again at night. A casual suggestion and a real decision coexist in the same day.
The model is not always “forgetting.” More often, it has experienced too many events and starts treating things from different times as the same present.
The nasty part of long context is that the model does not look ignorant. It looks over-informed by things that should not all count at once. Like a doctor on day three of a shift treating Monday’s warning, Tuesday’s patch, and Wednesday’s new request as the same patient note.
Compression is overnight, but yesterday becomes a class
What about context compression?
In this metaphor, compression is not “not sleeping.” Compression is closer to a real overnight transition.
Yesterday ends. Today begins.
But yesterday is not preserved in full. The agent harness records yesterday into a class and plays it for the new Ryland in the morning.
That class can be good or bad, long or short.
If it is good, Ryland is up to speed by 10 AM: what happened yesterday, why decisions were made, which traps to avoid, and where today should start.
If it is bad, Ryland spends the morning watching a messy surveillance tape. All the details are mashed together with no causality, no decisions, no priority. The class is not even over, and today is already afternoon.
If it is too short, key decisions disappear.
If it is too long, it is almost the same as not sleeping, just with a different flavor of exhaustion.
So the quality of compression determines what time Ryland actually starts working the next day.
This pairs naturally with SP-191 on how Claude Dreams cleans up the memory junk mountain for agents. Dreams is what happens when yesterday goes to sleep and gets reorganized; this essay asks whether tomorrow morning’s class was recorded well enough.
Good compression is not “make the summary as short as possible.” It is “let tomorrow’s Ryland stand in the right place with the least morning time wasted.” Bad compression is a surveillance transcript. Good compression is a nurse’s shift handoff: what happened, what was done, and what the next person must not screw up.
The agent harness decides which world Ryland lives in
Now zoom out. The agent harness is not a small wrapper.
It is the physics of Ryland’s world.
For the concrete harness side of the same idea, CP-226 on Natural-Language Agent Harnesses is the matching puzzle piece: change the world’s rules around the same model, and the model’s path through the task changes too.
The same model can feel like a totally different person depending on the harness around it.
Some agent harnesses are like spaceships.
Ryland wakes up with navigation, dashboards, files, tools, tests, and error alarms. The ship lets them read star maps, repair engines, inspect files, and run diagnostics. If something goes wrong, the ship reports it back so Ryland can correct course.
In that world, Ryland can travel through space.
Some agent harnesses are like mailrooms.
Ryland’s world is small: receive mail, read mail, write replies, send mail. Not many tools. Not a complex file system. No need to explore the universe.
That is not necessarily worse.
Mailroom Ryland has narrow abilities, but clear boundaries, higher safety, and a smaller blast radius. Spaceship Ryland can do bigger things, but they can also slip once and fly the ship straight into Mars. Ryland is an LLM, a probabilistic model. Hands get sweaty. Eventually, one of them slips.
So evaluating an agent is not just asking which model is inside. The same model inside a spaceship and inside a mailroom is basically two species. The model is Ryland. The harness is the universe.
Good agent workflow arranges the day
If the context window is a day, good agent workflow is not “stuff more things into context.”
It is arranging Ryland’s schedule.
The morning lesson should be short, sharp, and useful.
Daytime events should be ordered, not constantly dragging obsolete clues back into the present.
Important decisions should be written down when they happen, not reconstructed at midnight.
The overnight class should be recorded well, so tomorrow’s Ryland does not have to rewatch all the security footage.
The harness should know whether it is building a spaceship or a mailroom. Both are valid. What does not work is calling it a mailroom while handing Ryland a spaceship with no manual.
This is why “start a new chat” can be stronger than “keep reminding the model.”
The model is not necessarily getting dumber.
It is just too late in the day.
Before dawn
So do not think of the context window as only “how many words fit.”
It is Ryland’s day.
If the morning class runs too long, Ryland has less time to work. If daytime events are dumped in without order, morning, afternoon, and midnight start blending into the same present. If the overnight class is recorded badly, tomorrow’s Ryland begins the day inside a disaster recap.
Old models were koalas, awake for two hours. Modern models can stay awake for three days.
That does not make them gods.
It only gives humans a chance to arrange a decent schedule.
Do not ask whether Ryland can keep going a little longer.
Check what time it is first.