Low-latency co-speech gesture generation remains a critical bottleneck for interactive digital humans, as existing generative models suffer from high inference latency while retrieval-based methods lack semantic adaptability. To reconcile high-fidelity synthesis with instantaneous responsiveness, we introduce NeRAG, a Neuro-Explicit Retrieval-Augmented Generation framework. NeRAG decouples generation into two synergistic stages: a Transformer-based Neural Motion Planner that predicts semantic intent, and a graph-based Explicit Motion Executor that retrieves physically valid motion primitives. Crucially, we propose a Recurrent Motion State Feedback (RMSF) mechanism, which establishes a closed-loop system by feeding the actually executed kinematic state back into the planner, ensuring structurally enforced transition smoothness. Coupled with a Hierarchical Graph-Constrained Search algorithm, our system achieves 120 FPS throughput with an end-to-end latency of 150ms on consumer hardware, satisfying the HCI-defined threshold for perceived real-time interaction (<200ms). Extensive experiments demonstrate that NeRAG significantly outperforms state-of-the-art diffusion baselines in both motion quality and inference speed, enabling scalable deployment for LLM-driven avatars.