Knowledge has temperature
Why AI context setups rot in a month, and how to stop it.
AI context setups rot within a month. The reason is mechanical: every file gets treated the same way, when the files are not the same.
Look at what goes into a setup so an AI works like it knows you. Who you are. How you write. The decisions you’ve made. What you’re working on this week. The call you made an hour ago about pricing. All of it is context. None of it changes at the same speed.
Who you are barely moves. Write it once and it holds for a year. What you’re working on changes by the week. The decision you made this morning didn’t exist yesterday. That difference in speed is the whole thing. That’s temperature.
Cold, warm, hot
Cold files change yearly or never. Your identity, your writing voice, your principles. You set them down carefully and they sit there.
Warm files change monthly. The context around a project: who the client is, what the constraints are, where things stand. Real, but slow.
Hot files change weekly or faster. What you’re working on right now, what’s blocking you, what shipped yesterday. By next Friday half of it is wrong, and that’s fine, because you’ll have updated it.
One file sits on its own: the decision log. You add to it constantly, but you only ever add. You don’t rewrite a call you made three months ago. You append the new one underneath and let the old one stand as history.
Permissions follow temperature
How much you let the AI write to a file should depend on its temperature. This is the part most setups get wrong.
Cold files: the AI drafts, you approve, always. This is your identity. If the AI quietly edits how you sound, you won’t catch it in the moment. You’ll catch it three weeks later, when everything it writes is subtly off and you can’t say why. Identity drifts in silence, so you guard it.
Hot files: the AI writes directly, no approval step. The risk with a hot file runs the other way. If logging what you did today costs a round of approval, you won’t do it. The file goes stale, then you stop trusting it, then you stop using it. I’ve watched this happen, including in a setup of my own. Gate every file to be safe and the hot ones quietly stop getting updated, until the whole thing is frozen in the state it was in on day one. Friction is what kills hot files, so you take the friction out.
The log: the AI appends, never edits what’s already there. New decisions go in freely. Old ones stay untouched.
Why one rule for everything fails
Pick a single rule and apply it everywhere. Either the AI only writes to drafts you approve, or the AI edits whatever it likes. Both break, because they break in opposite directions.
Lock everything behind approval and the hot files calcify. The friction costs more than a wrong write ever would, so the updates stop and the setup freezes.
Let the AI write anything and the cold files corrupt. Your identity drifts a little each week, nobody signs off on it, and within a month the voice that made the setup yours is gone.
Match the permission to the rate of change. Guard what rarely moves. Free what moves constantly. Let the log grow and never rewrite it.
That’s why a setup built this way is still useful after the first month. The cold stays yours because you guard it. The hot stays current because nothing stops you updating it. The thing holds its shape.