Scratchpad
- Reasoning
- Standard term
A scratchpad is temporary working space an AI uses during a multi-step task, such as notes, partial results, or draft reasoning, kept separate from the final result it delivers.
Each word-piece a model writes draws only on what is already written. Writing partial results down gives later steps something to build on, so a problem too big for one jump gets solved in stages. An agent keeps a notes file so its findings survive when older parts of a long conversation are dropped.
Where you’ll meet it
You ask an agent to read forty customer interview transcripts and summarize the complaints. It starts a notes file in your folder, adding a line for each interview it reads on what that customer disliked. At interview twenty-six the session runs out of room and stops.
What to do
Start a new session, tell the agent to read the notes file first, and have it carry on from interview twenty-seven. Then compare two or three notes with their interviews, since the summary will likely be written from those lines. Delete the file once the summary is done.
Easy to mix up
Chain-of-thought is one kind of writing: reasoning laid out step by step. A scratchpad is the page that writing goes on, along with notes and partial results.
Check yourself
Planning a weekend trip, you open a chat assistant's thinking panel. One line says "museum closed Mondays, confirmed," and the final plan puts the museum on Tuesday. What would show you the closing day was looked up?
Go deeperA 2021 paper named it, and visible steps may leave causes out
Maxwell Nye and colleagues named the idea in "Show Your Work: Scratchpads for Intermediate Computation with Language Models" (2021). They trained models to write out the steps of tasks such as long addition and running short programs, which made them more accurate than answering directly.
Visible working can differ from how the answer formed. A 2023 study by Miles Turpin and colleagues found models changing their answers because of hints in the prompt while their written reasoning never mentioned the hints. Some products show a summary of hidden reasoning, which is a step further from what the model computed.
An agent's notes can also hold text copied from pages it read, planted instructions included, and on later steps the agent rereads them as its own words. A careful setup treats the notes file as untrusted input.