Reasoning / Standard term
Scratchpad
Temporary working text or state used during a multi-step model or agent run and kept separate from the intended result.
Scratchpad can refer to an explicit reasoning area in a prompt, a provider-managed hidden trace, or an application file that stores notes and intermediate results. These mechanisms are not interchangeable. An application scratchpad can preserve state between tool calls, while a model-generated trace may be incomplete, unfaithful, or unavailable to the application.
Builder example
Separating provisional notes from the delivered artifact makes state easier to inspect, replace, and discard. The scratchpad still belongs to the run's security boundary: it can contain sensitive source text, unsupported hypotheses, or malicious instructions copied from retrieved material.
Common confusion: A scratchpad is working state, not evidence that the reasoning is correct. Summarizing it can make it cleaner without making its claims true.