An organizational second brain for expert knowledge preservation
Large organizations often struggle to preserve the specialist knowledge that lives entirely inside their experts' heads.
3 min read·Thursday, 3 September
An abstract glass brain illustration representing the organizational second brain architecture that preserves specialist knowledge.
In many large organizations, the most valuable specialist knowledge lives inside people's heads rather than in written documents. In domains like compliance, experts often spend days on manual research or answering routine questions, leading to delays and inconsistent assessments.
To solve this challenge, engineers at Meta developed an AI agent that acts as a second brain. It captures how domain experts reason and makes that knowledge accessible to everyone across the organization. The system is designed to generalize to other enterprise fields such as finance, security, and engineering.
Off-the-shelf language models often lack the specific institutional context required for high-stakes decisions. General models can summarize general information, but they struggle to distinguish between what an organization could do and what it should do based on historical positions and company priorities.
The system relies on four distinct layers that depend on each other:
A structured knowledge system with a file structure that enables automated editing.
A reasoning layer with explicit procedures that make failure attribution tractable.
An evaluation framework that gates every change.
An improvement loop that feeds back into both knowledge and reasoning.
Organizing knowledge files
Instead of retrieving raw document chunks at inference time, which can be slow and inconsistent, the system uses a long-running offline process to distill source documents into structured knowledge files. These files are organized into a strict taxonomy of over 200 files.
Sample directory structure of the knowledge base showing position files and references. · Source ↗
Position files capture authoritative organizational stances, boundary conditions, and routing implications. Taxonomy and vocabulary files act as an authoritative glossary. Routing indexes match input characteristics to relevant positions, while gateway files define threshold tests the agent must pass before entering an analytical domain.
Composable reasoning recipes
Domain experts follow structured methodologies rather than simply recalling facts. The system uses composable procedures called recipes to guide the language model through multi-step analytical workflows. Knowledge files are declarative, meaning they state facts without prescribing procedures, while recipes are imperative and outline the analytical workflow.
This separation allows teams to add organizational positions or fix methodology flaws without editing the other layer. Restructuring the system into recipe-driven stages cut the tokens consumed per turn by around 80% compared to flat instruction files.
Keeping humans in control
Human experts retain authority over the final outcome through checkpoints and escalations. Checkpoints allow experts to review intermediate reasoning before the agent proceeds. Escalations occur when the agent encounters genuine ambiguity, handing the question over to the human expert.
Gemini Spark: Autonomous AI Agent Setup & Master BlueprintFree guide
Comments