Context window

A model reads a fixed amount of text each time it runs. That amount is its context window, and it is counted in tokens. Instructions, the conversation so far, and any retrieved material all have to fit inside it.

When the limit is reached, something must be dropped or left out. The model cannot use what it never receives.

• Agents that run long tasks fill the window quickly.
• Larger windows cost more per call and can dilute attention to the parts that matter.
• Choosing what to include is the core design problem.

• Recall returns results within a token budget that you set.
• You set the budget with level (low 8,000 tokens by default, medium 16,000, high 24,000, extra-high 32,000) or with maxTokens.

Not by itself. A larger window still has a limit, and the model reads everything inside it on every call. Harny recall returns results within the token budget you choose.