COMPACTION AND SUMMARIZATION

LEARNING GOALS Sooner or later, every long-running agent faces the same problem: the history grows until it threatens to overflow the window. In this chapter you learn the main cure, which is called compaction. You will learn what compaction is, how a token threshold triggers it, how to keep recent messages while summarizing older ones, how to make sure critical details survive, and how newer approaches let agents compress their own context actively.