02Context & memoryDeep dive ③

Pick what goes on the desk by 'what this turn needs', not 'what I have'

Relevance beats completeness — and what must be remembered lives outside the desk

Published2026-05-31
Read2 min read
TypeField notes
TL;DR

Choose what goes into context by "what this turn needs," not "what I happen to have." Relevance beats completeness. And anything that must survive across turns — decisions, constraints — gets stashed in a file outside the desk, because the desk gets wiped and re-laid once per turn.

Mid-way through a long task, you notice the agent drifting: it restates something already settled, proposes the thing you just ruled out. The reflex is to paste the entire history back so it "remembers." But pasting the whole history walks straight into the budget trap — overfilled, diluted, edges shedding. The right question isn't "what more do I add," it's "which exact thing do I re-lay."

Clearing the desk, in the end, is an act of choosing. And there are two principles for choosing.

01Relevance beats completeness

The "hand over everything" instinct optimizes the wrong thing. You don't need the agent to see everything — you need it to see the right thing for this turn. The function being edited, the constraint in force, the specific error in front of you: on the desk. The rest of the codebase, three decisions unrelated to this step, yesterday's logs: off it, brought in when you reach them.

One question does the job: "to get this turn right, what does it need to see — and only what?" The "only" is the hard part: cutting the important excess matters as much as supplying the right thing. Every spare item on the desk isn't neutral — it's a competitor for attention against the core.

02What must be remembered lives outside the desk

This is where "this turn's context" parts ways with "the whole task's memory." The desk is for this turn — and it gets cleared. So anything that must live across turns, left sitting on the desk, will eventually slide off: an architecture decision, a convention, a list of constraints.

The fix isn't to re-paste them every turn, but to persist them outside — a decisions file, a line of convention where the agent always reads, a short note. Then each turn you pull back only what that turn needs. You move from "hoping it still remembers" to "it can always read it, because I wrote it down."

These two principles close the cluster: the desk is for this turn (chosen by relevance, not completeness), while the whole task's memory lives off the desk (in a file, not in hope). An agent doesn't remember for you — it only re-reads exactly what you deliberately set in front of it.

End of pieceCluster 02 · 4/4
The author

craftagent is the notebook of someone still building — told over coffee, each story wrapped around a lesson paid for in full.