context window
The most tokens a model can hold at once: system prompt, conversation, files it read, and everything it has said so far. Frontier models in 2026 quote 200k to 1M tokens. When it fills, something has to be dropped or summarized, and the model has no memory of what left.
The window is not free memory. Every token in it is held in the KV cache, which sits in RAM next to the weights and grows as the conversation does. A large window is a memory cost you pay on your own machine and a token cost you pay per request in the cloud.
Filling it is also not the goal. Answer quality falls off well before the stated limit, and agent harnesses that re-send the whole conversation on every step pay for the entire window again on every tool call.