SYSTEM Cited by 1 source

Meta Organizational Second Brain

What it is

A domain-expert AI agent built by Meta that acts as a "secondary expert" for a specific compliance domain — making deep specialist knowledge available to anyone in the organization and preserving it durably. Unlike a typical domain-specific agent, its design integrates two layers (Source: sources/2026-09-02-meta-an-organizational-second-brain-building-an-ai-that-learns-from-experts):

1. A structured, auditable knowledge architecture that separates what the agent knows from how it reasons — see knowledge-reasoning-separation.
2. A self-improvement loop that compiles expert feedback into verified, regression-tested updates without model retraining — see self-improving-knowledge-base-loop.

Four layers

1. Knowledge system ("organizational second brain") — 200+ files in a strict taxonomy (Position, Taxonomy/Vocabulary, Routing index, Gateway), each declaring depends_on / referenced_by in YAML frontmatter to form a bidirectional dependency graph. Built by a long-running offline distillation process.
2. Reasoning layer — composable recipes (imperative multi-step procedures that reference knowledge but hold no facts), composed into pipelines with a top-level routing recipe. Enables progressive disclosure (~80% fewer tokens per turn vs the flat-instruction baseline).
3. Evaluation framework — gates every change via blind targeted replay + regression benchmarks.
4. Improvement loop — feeds corrections back into both knowledge and reasoning; enriches the regression suite on every landed fix.

The layers are interdependent: "Remove any one layer and the others degrade."

Knowledge vs RAG split

The system partitions sources by information density × usage frequency: high-density, frequently-referenced org reasoning goes in the curated wiki; sparse, situationally-relevant material is served via RAG.

Human control

Humans stay in control via checkpoints and escalations — the agent accelerates and structures work but does not replace expert judgment; every correction/escalation is also a training signal.

Results

• Outputs rated useful almost all the time after ~6 weeks / three sprints.
• Assessment time days → minutes.
• Zero regressions across improvement cycles; each fix strengthens the regression suite (regression-suite-enrichment-loop).
• ~80% per-turn token reduction from recipe-driven progressive disclosure.

Relationship to other Meta / wiki systems

• Sibling to Meta's AI Pre-Compute Engine (offline multi-agent extraction of compass-not-encyclopedia context files) — both bet on markdown-level encoded knowledge as a model-agnostic substrate; this system adds a rigorous feedback→verified-edit compilation loop on top.
• Related to Spotify's Vedder context layer for its data assistant.
• A production instance of Karpathy's LLM Wiki pattern (cited in the post), which is also the pattern this wiki itself follows.

Related

• knowledge-reasoning-separation, document-based-knowledge-editing-as-compilation, attribution-test-for-feedback
• composable-recipes, self-improving-knowledge-base-loop, independent-adversarial-review, deterministic-structural-linter
• companies/meta