What goes into the window, and how an agent remembers beyond it.

What you'll learn

A model knows only what is in its context window. Context engineering decides what goes in: which chunks of which documents, found how, cut where, packed in what order, what to drop or summarise when the window fills, and what to remember between sessions. Every claim here is measured, not asserted: on 6 articles of SQuAD and 200 of their questions, each with its answer's exact place in the text, so a retriever either finds the chunk that holds the answer or does not.

The animations are driven by Agent_Loop_Sim's context module. A small embedding model and a cross-encoder ran once, offline; their outputs ship with the site (int8 vectors and recorded scores), and the engine recomputes every chunk boundary, score, ranking and metric from them, identically in Python and in your browser. On this corpus BM25 finds the answer's chunk first 79% of the time, dense retrieval 71%, the two fused 81%, and fused then reranked 89%. No language model runs anywhere.