The Scratchpad Board: ReAct's Working Memory

From the Dongfang Yiteng execution agent (case by Bo Liang)

Definition

Reasoning gives the next step; action has to carry that step through to completion. An execution-type agent's reasoning advances only one step at a time, and after acting it decides what to do next from the result. So the action capability itself must carry both reasoning and execution: act one step, observe the result, and if there is still a gap to the goal, keep going, closing in on the goal step by step. This requires tool calls to be wrapped inside the action capability and scheduled by it, rather than the model calling tools bare. This is the ReAct action pattern, and it needs a piece of working memory to hold the execution record of the earlier steps. That piece of working memory is the scratchpad board.

How it works

Every round of ReAct contains an explicit trio of Thought, Action, and Observation: first decide what this step should do, then perform one actual action, then observe the result the action returns. The trio is appended as a block to the scratchpad structure. Round after round, the board accumulates a chain of blocks. Later steps do not start from zero; they read the board first and continue reasoning on top of the steps already completed: what the previous step found, what the previous tool call returned, all recorded on the board for this step to draw on. Its name comes from the shape of the structure — like a sheet of scratch paper that records the working line by line, to be looked back at when you need to see how far things have progressed.

The scratchpad board solves the model's amnesia problem. Every LLM call is stateless; the model itself does not remember what happened in the previous round. If each round's thought, action, and result is not explicitly recorded and fed back to the model, the agent may lose the key information found at the first step by the third step, and so repeat work already done, or make a wrong judgment on the basis of incomplete knowledge. The scratchpad board externalizes this execution trace into a readable, appendable data structure, giving a loop that was originally memoryless a continuous working memory, so each step can continue on top of the results of the previous steps.

When Dongfang Yiteng validated the quick-setup scenario for a payroll group, this board of ReAct's was the basis on which the action loop ran. After the user raised the need to set up a payroll group, the agent advanced step by step rather than planning out every step in advance: first it judged that a payroll-group template needed to be matched, performed the matching action, observed the result the matching returned, and appended this round's trio to the board; the next step read the board, learned that the template had been matched, and then judged that a snapshot needed to be created and afterward imported. The actual operation of each step is carried out inside the Orchestrator, and after it is done the progress is checked against the user's goal before deciding the next round. The control signal drives which specific action each step takes, while the board guarantees the continuous transfer of state across this sequence of actions.

When you need it

When an agent has to execute multiple steps in succession under one goal, and whether to continue and what to do next at each step depends on the real results of the earlier steps, you need a scratchpad board to record the execution trace. If there are clear dependencies between the steps and the macro order can be fixed in advance, ReAct hands control back to the task DAG and state machine, switching from step-by-step to plan-then-execute. The execution progress that the scratchpad board records and the narrative memory that ultimately gathers "what the user wants and how far it has been done" are two different kinds of state; the latter is carried by the memory envelope.

This is an ADPS blue-book concept. Back to Concepts or the pattern matrix.