Agent Memory and Context

1. The Problem: A Long Task in a Finite Window

An agent working a real task builds up history fast. Think of a two-hour debugging session, or a support case with fifty tool calls behind it. The context window is the model's working memory: the text it can actually read in a single call. That history grows faster than any context window can comfortably hold. And the agent pays quadratic-ish cost just to re-read that history at every single step