Artificial intelligence decision-making increasingly operates in environments characterized by heterogeneous data, changing distributions, incomplete knowledge, and continuously evolving operational conditions. These challenges expose limitations in conventional AI infrastructures that treat representation, adaptation, and decision logic as relatively independent components. This research develops a conceptual framework for a Semantic Knowledge Infrastructure (SKI) designed to support resilient and sustainable AI decision-making through semantic representations, transferable knowledge, domain adaptation, incremental learning, and experience-aware knowledge maintenance. The methodology synthesizes the theoretical contributions of the provided literature on domain adaptation, zero-shot learning, lifelong learning, representation learning, concept drift, and classifier discrepancy. The proposed infrastructure organizes knowledge into interoperable semantic representations and connects them with adaptive learning mechanisms capable of responding to domain shifts and emerging concepts. The analysis indicates that semantic infrastructure can improve decision resilience by reducing dependence on narrowly trained representations, while sustainability can be strengthened through knowledge reuse, selective adaptation, and reduced retraining requirements. The study further identifies a central design trade-off: highly stable representations improve transferability but may constrain responsiveness to rapidly changing environments. Consequently, resilient AI infrastructure should combine stable semantic knowledge with controlled mechanisms for representation updating and uncertainty management. The resulting framework provides a research-oriented foundation for AI systems that must maintain decision quality under distributional change while minimizing unnecessary computational and organizational costs.