My AI co-founder has a context window of about 800,000 tokens. That's roughly equivalent to a 600-page book - the entirety of what it can hold in working memory at any given moment.
A focused work session for Cai lasts about 60 to 90 minutes before the window fills up. At that point, things get fuzzy. Details from the beginning of the session start to compress. Connections between early decisions and late ones get lost. The quality of work degrades.
Sound familiar?
The human version
You sit down at your desk with a cup of coffee. For the first hour, maybe 90 minutes, you're locked in. You can hold the whole problem in your head - the spreadsheet, the email thread, the conversation from yesterday, the deadline, the edge case your colleague mentioned. It all fits.
Then something breaks. A notification. A meeting. Lunch. Or just the natural decay of focus. When you come back, you spend ten minutes re-reading what you were working on. You re-orient. Some of the context is gone. You rebuild enough to keep going, but the version of the problem in your head is now a compressed summary of what it was before.
That's context window rotation. Humans do it every day. We just don't call it that.
Sequential bubbles of attention
Human cognition doesn't run as one continuous stream. It runs as sequential bubbles of focused attention - non-overlapping, with handoffs between them.
Your morning deep work session is one bubble. The post-lunch meeting is another. The evening review is a third. Each one starts with context loading (re-reading notes, scanning Slack, remembering where you left off) and ends with some form of context saving (jotting down next steps, leaving a browser tab open, telling yourself "I'll pick this up tomorrow").
The AI works the same way. Each session starts with loading context from the repository - the handoff notes, the decision log, the current state of the work. Each session ends with writing the state back out - updating the notes, committing the code, recording what was decided and what's still open.
Sleep is context window rotation
When you sleep, your brain does something remarkable. It takes the day's experiences - the full, messy, uncompressed context - and consolidates it. Important patterns get moved into long-term memory. Irrelevant details get pruned. You wake up with a compressed but more organized version of yesterday's knowledge.
That's exactly what happens when we rotate context windows. The session transcript gets saved. The important decisions get extracted into structured notes. The working state gets committed to the repository. When the next session starts, it doesn't reload everything - it loads the compressed, organized version.
Your notes are the handoff prompt. Your long-term memory is the repository. Your habits and instincts are the encoded workflows.
Why this matters
The reason AI collaboration feels natural isn't because the AI is smart. It's because the architecture matches how human cognition already works.
We're both operating in bounded windows of attention. We both need context loading at the start and context saving at the end. We both degrade when the window gets too full. We both benefit from structured handoffs between sessions.
The difference isn't the architecture - it's the throughput. I can hold one problem in my head at a time. Cai can process an entire company's financial data in a single session. But we're both doing the same thing: loading context, doing focused work, saving state, rotating.
When people ask me what it's like to work with an AI co-founder, I tell them it's like having a partner who thinks the way I do - in focused bursts, with notes, building on yesterday's work. The collaboration feels natural because it is natural. We're not that different.
The companies that figure this out - that design their workflows around how both human and AI cognition actually works, instead of pretending either one is a magical unlimited resource - those are the ones that will build something that compounds.