Lesson 3 of 10 · AI Agents and Automation That Finish Work
In this lesson. Store only what the next step needs, then pick the work back up.
What you will learn
- Scratch versus durable memory
- Checkpoints for long jobs
- What to reload after a crash
Walkthrough
Agents that remember everything get slower and more confused. You will separate scratch notes, durable facts, and transcripts. For long jobs you will checkpoint progress so a crash does not restart from zero. The design question is always: what must be true when this process wakes up tomorrow?
Work through the ideas in order. After each point, pause and connect it to a task you already do — a document, a workflow, or a feature you own. The goal of AI Agents and Automation That Finish Work is usable skill, not a pile of notes.
If something is unclear, rewrite it in your own words before you continue. Teaching the step back to yourself is the fastest way to see gaps.
Practice
Write the memory fields for a multi-hour task and the checkpoint you would save every 10 minutes.
Keep the first attempt small. A finished example you can reuse beats a perfect plan you never run.
Check your understanding
- Can you explain the goal of this lesson in one sentence to a teammate?
- Where would you apply “Scratch versus durable memory” in your own work this week?
- What would you change on a second pass of the practice?
Next. Continue to the following lesson when the practice has a real artifact, even a rough one.