Context window
Also: context length
A hard limit on how much text the model sees in one forward pass, set by its training and positional encoding and bounded in practice by attention cost and KV-cache memory. Anything beyond it must be truncated, summarized, or retrieved back in.
Example
With a 128K-token window, a ~100K-token book fits with ~28K tokens left for instructions and the answer; a long chat that exceeds the window must drop or summarize its oldest turns.