Reasoning / Standard term
Scratchpad
A workspace where a model writes down intermediate thoughts, calculations, or working notes before producing its final answer. The AI equivalent of scrap paper.
When you solve a long division problem, you jot partial results in the margin. A scratchpad serves the same purpose for an AI model: a section of the prompt labeled "think here before answering," a hidden reasoning buffer the provider manages internally, or a temporary file an agent writes to while working through a multi-step task. It lets the model hold intermediate state that would otherwise exceed what it can track in a single generation pass.
Builder example
Scratchpads separate working state from user-facing output. When building an agent that researches a topic across multiple sources, you want the agent to accumulate notes in a scratchpad and present a clean synthesis to the user. Without this separation, internal reasoning artifacts like dead-end hypotheses, raw data fragments, and self-corrections leak into the final response and confuse users.
Common confusion: Scratchpad content can contain hallucinations, sensitive data from earlier context, or misleading intermediate reasoning. Exposing raw scratchpad output to users is a common source of trust and safety issues.