We rethink parametric retrieval-augmented generation with DuoPRAG, which integrates knowledge and reasoning capabilities through rank-disentangled and gated Low-Rank Adaptation (LoRA) composition. Existing parametric RAG methods encode document knowledge into LoRA parameters while leaving reasoning to be supplied through lexical chain-of-thought (CoT) prompting. This introduces three failure modes in outputs and contradicts the prompt-independent inference that parametric RAG aims to achieve. DuoPRAG parameterizes knowledge and reasoning into disentangled LoRA adapters, fused through a learnable mask and rank-specific gates. We achieve state-of-the-art performance on multi-hop QA benchmarks, outperforming the strongest parametric RAG baseline while preserving the parametric inference framework. Our results demonstrate that jointly parameterizing knowledge and reasoning is a natural and effective extension of the parametric RAG paradigm. Our code is available at https://github.com/naknak-choi/DuoPRAG.