Scratchpad
A scratchpad is the ephemeral working memory space in an agent's context window where intermediate reasoning steps and partial results are written and read during a single task.
title: "Scratchpad" type: glossary tags: [#reasoning, #planning] created: 2025-07-14 updated: 2025-07-14 status: stub
Scratchpad
A scratchpad is an intermediate working memory space where an AI agent writes and reads partial reasoning steps, calculations, or intermediate results before producing a final output.
Overview
The scratchpad metaphor captures how agentic AI systems — particularly those using chain-of-thought or ReAct Framework patterns — externalize intermediate reasoning within the context window. Rather than producing a final answer directly, the agent generates a sequence of intermediate thoughts, tentative conclusions, or working notes that it can reference and revise before committing to an action or output.
In practice, a scratchpad is not a separate system component but a convention for how reasoning traces are structured in the agent's context. The "Thought:" prefix in ReAct steps serves as a scratchpad: the agent is effectively writing notes to itself that it will read when generating the next step. This mechanism is what allows agents to work through complex multi-step problems that would be difficult to solve in a single forward pass.
Scratchpads are closely related to Memory Systems but are distinct: scratchpad content is ephemeral, existing only within a single task's context window, while long-term and episodic memory persist across sessions. The scratchpad is best understood as the agent's working memory for a specific task instance.
Related Pages
Used by: ReAct Framework, Agentic AI See also: Memory Systems, Agent Architecture
Page type: glossary | Status: stub