Context Window
The maximum amount of text (measured in tokens) a model can consider at once — its working memory. Everything counts against it: your instructions, examples, conversation history, and the reply itself. Modern LLMs range from around 128K tokens to millions, but even huge windows have a catch: models attend less reliably to the middle of very long contexts. Practical impact: when a long chat seems to forget early instructions, you have overflowed or diluted the window. For big creative projects, restate critical constraints in fresh messages instead of trusting a 200-message history.