Scratchpad
A space where the model writes intermediate thoughts and notes to itself while working through a problem.
Think of it like
The margin of an exam where you scribble your working before writing the final answer in the box.
Example
Before answering a logic puzzle, the model jots the constraints and rules out options in a scratchpad, then gives the clean answer.
How it actually works
Letting a model externalize its reasoning into text it can re-read improves hard, multi-step problems — it is chain-of-thought given a place to live. In agents, the scratchpad also holds notes across loop turns: what’s been tried, what’s left. Sometimes it is shown to the user, sometimes hidden; either way it is scaffolding for thinking, not the deliverable.
For product teams
Why “show your work” prompting makes models measurably more correct.
For engineers
A region of context the model writes intermediate reasoning or state into, re-read on subsequent steps.
Related
- Chain of thought — The reasoning style a scratchpad enables.
- Agent Memory — Longer-lived notes that outlast a single scratchpad.
- ReAct — Where it fits in the think-act rhythm.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome