Context Window
How much text the model can see at one time.
Language models have no memory between calls. What looks like a continuous conversation is the whole history being resent on every turn, from the top. The context window is the limit on that.
Two practical consequences follow. The first is cost: what you pay for is the number of turns, not the size of the task. The second is memory: anything that matters and is not written to a file disappears when the window fills or the session ends.
Hence the asymmetry worth internalising: what sits in the window is volatile and expensive; what sits in a file is durable and cheap.
Read the article: I counted dcode's loop and found the stopping condition