Abstract We present HATI-EBM, a reference architecture for enforcing alignment and agency-preservation constraints in generative AI systems through continuous, physics-like system dynamics rather than symbolic policy layers. The architecture places a lightweight Energy-Based Model (EBM) beneath a generative model as a Safety Governor: a narrow constraint evaluator that assigns energy scores to system states and interaction trajectories, dampening or rejecting generations that violate defined safety boundaries. We argue that real-time constraint enforcement requires hardware–software co-design, and demonstrate that Unified Memory Architecture (UMA) with high memory bandwidth (≥800 GB/s class) is the critical enabling property. We formalise two core mechanisms: the Agency Transfer Gradient (ATG), a computable measure of increasing dependency asymmetry in human–AI interaction, and the Energy Wall, a thresholded regime in which generation is progressively dampened as system state approaches unsafe regions of the constraint manifold. Prototype evaluation on consumer-grade UMA hardware shows that EBM-based safety governance adds 3–8% inference latency overhead while reducing constraint violations by 40–60% compared to post-hoc symbolic filters, with substantially lower false-positive rates. All results are clearly labelled as prototype-level. The architecture is designed for local deployment, requiring no centralised governance infrastructure. Keywords: energy-based models, safety governance, unified memory architecture, memory bandwidth, agency preservation, hardware–software co-design, local AI deployment