CONTEXT AS THE SCARCE RESOURCE

THE COMPUTER MEMORY COMPARISON To understand why this matters, compare the context window to Random Access Memory (RAM) in an ordinary computer. RAM is the fast, temporary workspace a computer uses for whatever programs are currently open. When you open a web browser, a spreadsheet, and a music player, each one takes a slice of RAM. RAM is fast, but it is limited, and it is cleared when the computer is turned off. If you open too many programs, the computer slows down, or it must close something to make room. Now map this to an agent. The context window is the agent's RAM. It is fast, in the sense that the model can look at anything inside it instantly while thinking. It is limited, in the sense that only so much fits. And it is temporary, in the sense that when a call ends, nothing is automatically kept. Just as a computer with limited RAM must close some programs to open new ones, an agent with a limited context window must drop or summarize old messages to make room for new information. That sentence is the heart of this entire Part. Everything else, including compaction, memory tiers, and context builders, is a technique for doing that dropping and summarizing wisely instead of carelessly.