Procedural Memory
Re-deriving 'how we file a Jira bug' every time wastes tokens and drifts from team process.
Riding a bike: you do not re-read the manual; you run the skill.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Overview
Scroll inside the frame to explore · use + / − to zoom up to 200%
Scroll inside the frame to explore · use + / − to zoom up to 200%
Reuse a skill
Scroll inside the frame to explore · use + / − to zoom up to 200%
Match the task to a playbook, then run the saved steps.
Key Takeaways
- 1.Procedural memory is how-to: skills, playbooks, and saved procedures. It is 'run these steps', not 'this fact is true'.
- 2.Store as prompts, graphs, or tool sequences the agent can reuse. Version playbooks — a bad saved procedure will repeat forever.
- 3.Procedural memory is reusable how-to. Keep it as a named procedure with inputs.
- 4.Do not hide it only inside a long system prompt you cannot version.
Learn elsewhere
- →Tool Calling — Phase 7
- →Context Engineering — Phase 6
- →Agent Design Patterns — Phase 17
Real Example
Scenario
Skill `file_jira_bug`: collect repro, severity, component, then call the Jira tool. The agent loads that skill instead of inventing a new process.
What you would do
In Agent Memory, apply Procedural Memory to this scenario: Skill `file_jira_bug`: collect repro, severity, component, then call the Jira tool. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Commands
Commands to Remember
How-to, not factsNamed playbooksVersion and testRetire bad skills
Cheat Sheet
Quick recap
quick ref- •Bike skill, not wiki
- •Saved sequences
- •Version playbooks
- •Different from facts
Common Mistakes
- ✕Stuffing all how-to into one giant system prompt
- ✕Never versioning a procedure after it fails
- ✕Confusing a fact ('we use Jira') with a procedure ('how to file')