Written by Martin Ågren - October 25, 2026
Context Window
Context window = Maximum context (tokens) a LM can consider at once. Includes: System prompt + chat history + prompt + output + etc.
Tokens = Units of text (parts of words / words / symbols) that a language model interprets and/or generates.
Limit passed = If limit of context window is reached: LM forgets beginning of chat OR summarizes start internally OR "400" error (for API)
Context rot = Performance decline (of a LM) that can occur as context grows. Can occur before context window limit is reached.
Frontier cloud model = Usually very large context window - But short conversations are still often preferable to avoid context rot.
Local model = Model that you run on your computer = Can have small context window. In that case, keeping chats short is nessesary.