Context windows: LLMs and human brains
How LLMs made me notice my own limits, and why I am more careful now with focus, writing, and mental context.
We often underestimate what LLMs are. The more I use them, the more I see parts of myself in them. One of those parts is the context window.
I have never cared more about my mental integrity and focus than recently. We are in an era where it feels like we can build or solve almost anything. That is exciting, but it also creates pressure. I often feel like I am not doing enough, not moving fast enough, not taking enough advantage of the tools in front of me.
This reflection boosted my ego in a strange way. I do not like pointless things, and I like being rewarded for my brain. Some days I honestly think my brain is the best thing I own. Then, after a few long coding sessions, I get cooked. The machine slows down. The sharpness is still there, but the active state is gone.
The window fills
That is the context window. Not in the exact technical sense, but close enough to be useful. A brain can hold only so much live context before older details get compressed, distorted, or dropped.
LLMs have compaction. They can summarize a long conversation and keep going with a smaller version of the state. We do something similar, but it is messier. We call it memory, notes, intuition, experience, or "I think I remember why I did that." Sometimes that works. Sometimes it very much does not.
Project switching is not free
This is why switching between projects is expensive. It is also why switching between features inside the same project is expensive. A feature is not only the ticket title. It is the branch, the files, the meeting, the bug that only appears after a weird path through the UI, the option you rejected, the reason you rejected it, and the small emotional memory of what felt wrong.
When you leave that and come back later, you are not just opening a repo. You are reconstructing a mental world. If you do that enough times in a day, of course you feel tired. Of course the work feels slower than it should.
Going old style
What I do often is go old style: I draw or write my thoughts. I recently bought a notebook without lines so I can write wherever I want on the page, lose myself a bit, and put down the context in the shape it naturally has in my head.
There is something different about writing by hand. You have to materialize the thought. You cannot just keep it floating as a vague feeling. You have to decide what it is, draw the relationship, name the constraint, and make it real enough that your hand can put it on paper.
That gives me time to think. It slows the loop down in a good way. The goal is not to be aesthetic or organized. The goal is to protect the context window before it collapses under too much active state.
Taking care of the machine
I think this is going to matter more, not less. Tools are getting stronger. The ceiling is higher. It is easier to start five things, explore three possible solutions, ask an agent to implement one path, review another, and still believe you should have done more by lunch.
But the human context window is still human. It needs care. It needs breaks. It needs external memory. It needs fewer pointless switches. It needs the humility to admit that "I can do this" and "I can keep all of this in my head forever" are not the same sentence.
We are amazing machines. We should take care of them.
For me, sometimes that looks like a notebook with no lines, a pen, and enough patience to let the thought become real before I ask my brain to carry the next one.