1. The Mental Model

The framing that makes everything else click: the context window is RAM, not disk.

RAM is fast, expensive, and finite. You don't load everything you own into RAM before running a program. You load exactly what the program needs, right when it needs it. The same applies to LLM context: every token you put there displaces something else, costs money, and competes for the model's attention.

This reframes the engineering challenge. It's not "how do I give the model more information?" but "what is the minimum viable set of information the model needs to succeed?" Every technique in this page is an answer to that second question.

The parallel with system architecture holds further. A CPU without good memory management stalls. An LLM without good context management hallucinates, loses coherence, and drifts toward generic outputs. Optimizing context is a reliability investment, not merely a cost-cutting exercise.