Humans think with notebooks, whiteboards, and to-do lists because cognition is worse when everything must be held in the head. Agents are no different: externalized state — scratchpads, plan files, structured notes — reliably outperforms the same reasoning confined to the context window.

The mechanism is simple and measurable. A context window is a scarce, attention-diluted, pay-per-token resource; files and notes are effectively free, persistent, and addressable. An agent that writes its plan to a file can consult step 1 after 40 steps of drift; an agent keeping the plan in-context watches it fade under later observations. The Scratchpad technique showed the benefit for intermediate arithmetic as early as 2022; every coding agent since — from SWE-agent's interface design to CLI agents that keep NOTES.md — is externalized cognition as product surface.

The design space has three recurring artifacts. The plan file: the current task graph, explicitly versioned, updated as steps complete — the agent's contract with itself. The scratchpad: working calculations and hypotheses, deliberately disposable. The knowledge notes: durable facts worth carrying across runs — the write path into long-term memory. Conflating them is the common mistake; a scratchpad that survives into next session is stale bait, and a plan kept only in context is a wish.

Disciplines That Make It Work

• Write after reasoning, not instead of it: the note records a concluded thought; stream-of-consciousness files are noise with a filesystem.
• Re-read at decision points: consult the plan before each action; the file only helps if reading it is part of the loop.
• Expire deliberately: scratchpads reset per run; plans persist per task; knowledge notes persist per project — mismatched lifetimes cause subtle stale-state bugs.
• Review the residue: post-run, the files are the audit trail — what the agent believed, planned, and learned, readable without replaying tokens.

There is a deeper claim here, going back to distributed cognition in cognitive science: intelligence is a property of the agent-plus-artifacts system, not the agent alone. The engineering takeaway is refreshingly concrete — before buying a bigger context window, buy your agent a notebook.