Procedural memory

Procedural memory is the agent's how-to knowledge. Not facts, not episodes, skills. "How I usually handle this kind of task." "The recipe that worked last time." Agents that build up procedural memory get better at recurring tasks over time.

What qualifies

• Step-by-step recipes for recurring tasks
• Default parameters that worked in prior cases
• Decision rules: "if X, usually do Y"
• Templates for common outputs

How it's built

After a successful task, agent (or human) distills the solution into a reusable recipe:

{
  "procedure_id": "deploy_staging",
  "description": "How to deploy a branch to staging",
  "steps": [
    "Confirm branch is up to date with main",
    "Run tests locally",
    "Push to remote",
    "Trigger staging deploy workflow",
    "Verify at staging.example.com"
  ],
  "constraints": "Don't deploy on Fridays after 3pm"
}

How it's used

When agent starts a task matching an existing procedure, retrieve it and follow (or adapt). Much faster than reasoning from scratch.

The learning loop

1. Agent attempts new task, succeeds after some exploration
2. Post-task: distill successful trajectory into a procedure
3. Next similar task: retrieve the procedure, apply directly
4. Refine procedure over time as edge cases emerge

Relationship to other memory types

• Short-term = current working memory
• Long-term = facts that persist
• Episodic = specific past events
• Procedural = reusable how-to

A mature agent system has all four, layered.