AGENT ATLASa field guide to agent workflows
Back to Memory & context
Memory & context

Working memory

Let the agent keep notes within a task.

1 agentCore

How it works

  1. 1Give the agent a structured place to write notes and state as it works.
  2. 2Have it update the scratchpad each step — plan, findings, open questions.
  3. 3Feed the scratchpad back into each subsequent step.
  4. 4Clear it when the task ends.

Use it when

Multi-step tasks where the agent needs to track state, intermediate results, or a to-do list across steps in one run.

Reach for something else when

Short single-step tasks — the bookkeeping just adds noise.

Where you stay in the loop

The scratchpad is the agent's, but it is also your window into its reasoning. Read it to see the plan and state — it makes a single run legible and correctable.

In the wild

A research agent maintains a running "findings so far + still need" list it updates after each search.

Hand this to your agent

Maintain a working-memory scratchpad as you do this task.

Keep a structured block with: GOAL, DONE, FINDINGS, OPEN QUESTIONS, NEXT.
Update and re-print the whole block after every step before continuing, so
your state is always explicit.

Task: <...>

Replace the <…> placeholders, paste it into your agent, and it'll scaffold the workflow with you.