It is measured in tokens, which are pieces of words, and it covers the prompt, any documents or conversation history and the reply together. Once a conversation exceeds it, older material must be dropped or summarized, because the model cannot see it.
Windows have grown from a few thousand tokens in early models to hundreds of thousands, and in some models a million or more. A larger window lets a model take in a whole document or codebase at once.
A big window does not guarantee that the model uses everything well. Models can miss details placed in the middle of a long input, and cost and speed worsen as the input gets longer.
Check the real limit, test retrieval of facts at different positions, send only what is needed and consider retrieval methods when the material is far larger than the window.