Working definition

Context window means the model retains only within a finite window — attach project files so “it knew” is not chat folklore.

How it works

You trust the thread because last week the model “remembered.” Then it forgets, then it believes it knows — shortcuts, confident amnesia, garbage in → garbage out. The Club move is operational: put mission, allowed/forbidden, and expected output in markdown (skills as contracts) and attach them so memory is not theatre (AI Orchestrator · Quality of Questions).

Wire to Critical thinking on AI: challenge answers and challenge the fantasy of infinite memory. Pair with Tripartite reins: you own the file layer; the model owns labour inside the window. Probabilistic vs deterministic prompting (temperature) still sits on top of whatever you actually put in context.

Distinction from Agent orchestration: orchestration names roles; context window names the memory physics those roles run on. Distinction from System Within the System: the system is your Personal OS container; the context window is the model’s hard limit inside that system.

Signature

The image is the score left on the stand — the orchestra plays only what is on the page in front of it. Room: 2026-08-07attach markdown project files so memory is not chat theatre.

The WOW: it knew → forgot → believes it knows.

Your move this week

  1. Pick one recurring AI collaboration (podcast notes · outreach · research).
  2. Write or refresh one short project file: mission · allowed/forbidden · expected output.
  3. Attach it on the next run — then challenge one answer in three before you ship.

Scientific grounding

Cite

Long-context models do not use the full window uniformly — middle content is easier to miss (Liu et al.). Finite capacity is older physics than transformers (Cowan). Club doctrine: do not trust chat folklore; pin what must persist in attached files.

Adjacent

Be Your Real Boss