Externalize your memory

This lesson exists because of how it got written. The session that produced this whole Field Manual ran long. It went through a security pass, a design rabbit hole, a newsletter decision, a dozen calls made and remade. A normal chat thread would have hit its limit and started forgetting the early half by the time we reached the end. This one didn't lose a thing, and the reason is the entire point of this lesson: none of the important state lived in the thread. It lived in a memory the agent reads at the start of every session and writes to as we go. Coming back to the work isn't picking up a cold conversation. It's loading a save point.

Here's the fact underneath it, and it's not optional or a matter of taste: a context window is finite, and it degrades as it fills. Every model, all of them, gets less sharp as the conversation gets longer, because there's simply more to hold and a hard ceiling on holding it. Everyone hits this wall. The thing almost nobody does is build for it. The move is to stop treating the thread as the place your work lives, and start treating it as disposable, because your real state is written down somewhere permanent.