From Amnesia to Expertise.
A visual exploration of persistent mental models in AI agents.
SCROLL TO BEGIN
Generic agents are brilliant but forgetful. They burn tokens re-learning context every session.
Try it: Tell the agent your project name is "Orion". Then restart the session and ask "What is my project?"
To create an expert, we externalize memory into a structured file (YAML/Markdown). The agent consults this file before answering.
Lab: Edit the YAML file directly. Change users to clients. Then ask the agent: "What table stores people?"
True experts validate their models. If the code (Source of Truth) changes, the expert detects the drift and updates its Mental Model automatically.
CREATE TABLE users (
id INT,
email TEXT
);
table_name: users
We don't write experts manually. We use a Meta-Agent to spawn specialized experts and generate their initial mental models.
Lab: Define a domain (e.g., "WebSocket", "Payment") to generate a new expert.
The Agent Expert accumulates wisdom. While generic agents stay flat, the Expert's utility grows over time.
STATUS: EXPERTISE ARCHIVED.