The context window is the amount of text the model can take into account at once before answering.
Imagine a conversation partner reading the chat through a narrow frame. The frame fits the latest messages, the instruction and a piece of a document. Everything left outside the edge doesn’t exist for them right now — even if they saw it a minute ago.
At work this shows when AI confidently changes a function but breaks a rule from the beginning of the conversation. Or when an agent gets a large repository and “forgets” an important file: not because it’s lazy, but because that file didn’t make it into the current frame.
So a long chat isn’t a knowledge base. Important constraints are better repeated in the task, stored separately, and only the needed pieces of code and documentation passed in.
The main limitation: more context isn’t always better. Extra text dilutes the model’s attention the same way a long spec dilutes a developer’s attention. Context shouldn’t be accumulated but curated.