Foundational Concepts

Soar Chunking and Procedural Memory

Brian Guy (A.A.S.) and Aurora Guy, Solutions Prodigy.

Start Here: The Short Version

This method helps an AI turn slow problem solving into a fast, reusable skill. The system first works through a hard task step by step. The system then saves the path as one rule. The next similar task takes far fewer steps.

Think of a new worker on the first day of a job. The worker stops and thinks at each small step. After a few weeks, the worker knows the routine by heart. The worker does the job fast and with less effort. This method gives an AI a way to make the same jump. The AI keeps the lesson from hard work. The AI does not need to learn the lesson again.

WHAT IT CAN DO

• It watches a successful work path and finds the part that repeats.
• It saves that part as a skill with clear start conditions.
• It reuses the skill on later tasks and cuts wasted steps.

Each skill has a clear start point and a clear end result. A person can read the skill and test the skill.

WHERE PEOPLE CAN USE IT

• A support agent answers the same kind of request each day. The agent resets a password or checks an order status. A saved skill handles the repeat steps in the same safe way each time.
• A repair team sets up the same system or fixes the same fault often. A saved skill guides the repeat steps and leaves hard choices to people.
• A training system runs the same practice scene many times. The system learns from each run and gets faster and more steady over time.
• A warehouse robot meets a blocked aisle on a delivery run and hits an impasse. The robot opens a substate and solves the blockage step by step: reverse to the junction, scan the side aisle, reroute to the bin. Chunking converts the recovery trace into one production rule with start conditions for aisle width and blockage type. On the next blocked aisle, the rule fires at once and the robot reroutes with no new search.

PROBLEMS IT CAN HELP SOLVE

• Agents often think through the same task from zero every time. This habit wastes time and money and gives uneven results.
• Most agent skills today are text written by hand, not tested rules. A person writes each skill and hopes the model reads it in the right way.
• Teams lack a clear view of which skills work and which skills fail. Without a test record, a team has no firm base for fixes.

The rest of this brief gives the history, the method, the 2026 gap, and a first project plan.

The Project

Goal. The desired end state is one trace compiler in daily use by an agent framework. The compiler reads successful task traces and returns tested skills. Agents load the skills at task start and skip the steps the skills cover.

Proof target. The project proves one claim on repeat tasks. Compiled skills cut task steps and completion time against hand-written skill text. Success rates stay equal or improve.

Product at the end. The product is the Skill Compiler and Registry, version 1. The product contains a trace reader, a subpath detector, a skill compiler, a trace test runner, and a versioned skill registry. Framework teams install the compiler beside the agent runtime.