What is procedural memory in AI agents?
Procedural memory is an agent's stored knowledge of how to do a task, as opposed to facts or past events. How it is learned from verified runs.
Updated
Procedural memory in an AI agent is stored knowledge of how to do something: the steps, tools and conditions for a kind of task, as opposed to facts (semantic memory) or past events (episodic memory).
The three kinds of agent memory
The terms come from cognitive science and are now common in agent design:
- Semantic memory: facts and knowledge. "Refunds above $500 need approval."
- Episodic memory: specific past experiences. "On Tuesday, the refund for order 4821 failed with a 403."
- Procedural memory: how to act. "To refund a large order: request approval, then retry the refund with the approval ID."
Procedural memory is what lets an agent stop re-planning tasks it has done many times.
Learning procedures from episodes
Procedural memory is usually distilled from episodic memory: many runs of the same kind of task are compared, and the steps that consistently led to success become a procedure. Doing this well needs verified outcomes (otherwise failed runs teach), merging of differently worded tasks, and conditions for steps that apply only sometimes.
In Retrack
Retrack's playbooks are procedural memory learned from verified runs, with conditional steps and pitfalls. See Playbooks and the agent memory guide.