The maximum amount of text, measured in tokens, a model can take into account at once.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
The context window is the model's working memory: instructions, documents, and the conversation so far must all fit inside it, and anything outside it simply does not exist for the model. It is measured in tokens because that is what the model reads.
Window sizes have grown from a few thousand tokens to millions, but bigger windows cost more compute per call (attention relates everything to everything), and models use the middle of very long contexts less reliably than the edges. The window's limits are why retrieval techniques like RAG exist.