USD Cognitive Substrate: A Deterministic Architecture for Adaptive AI State Management This specification proposes applying Pixar's Universal Scene Description (USD) composition semantics as an architectural framework for managing cognitive state in LLM-based applications. Rather than treating USD as a container for 3D scene data, this work reinterprets its core mechanisms—LIVRPS composition ordering, opinion strength resolution, non-destructive layer stacking, and variant set selection—as primitives for organizing and persisting AI agent state. The architecture addresses key challenges in LLM agent systems: state persistence across sessions, deterministic behavior reproduction, explicit reasoning transparency, and graceful state recovery. By mapping cognitive constructs (memory, context, goals, persona) to USD primitives (prims, attributes, layers, variants), the framework inherits USD's battle-tested composition engine and tooling ecosystem. Key contributions include: (1) formal mapping between USD composition arcs and cognitive state operations; (2) a layered state model enabling non-destructive cognitive modifications; (3) variant-based persona and context switching with deterministic resolution; (4) integration patterns for production AI pipelines. This work emerges from 16 years of VFX production experience with USD pipelines and ongoing development of AI-integrated creative tools. The specification targets researchers and practitioners building persistent, inspectable, and reproducible AI agent systems.
Joseph O. Ibrahim (Tue,) studied this question.