An Organizational Second Brain: Building an AI That Learns from Experts (engineering.fb.com)

A new AI agent has been developed to serve as a secondary expert within organizations, efficiently capturing and preserving deep specialist knowledge. This innovative system integrates a structured knowledge architecture that distinguishes what the agent knows from the reasoning process employed, alongside a self-improvement mechanism that compiles expert feedback into permanent, regression-tested updates without the need for model retraining. This is particularly significant for large organizations struggling to systematically capture tacit knowledge that exists in the minds of their subject matter experts. By effectively transforming expert insights into an institutional memory, the AI helps streamline processes, allowing experts at Meta and similar organizations to focus on complex problem-solving rather than routine inquiries. The agent operates on a four-layer architecture, which includes knowledge files, routing indexes, and composable procedures (or "recipes") that guide the reasoning process. This design enables precise control over how data is accessed and used, ensuring that the expertise reflects the organization's unique context. Additionally, the system's self-improvement feature allows it to learn continuously from human feedback, enhancing its ability to provide accurate and relevant insights without constant retraining. This approach not only boosts efficiency but also fosters a culture of shared knowledge and collaboration, making the organization's specialized expertise more accessible and effective across various domains like compliance, finance, and engineering.