Urban Digital Twins (UDTs) have emerged as strategic instruments for urban planning and governance, generating unprecedented volumes of heterogeneous knowledge assets — policy documents, municipal acts, citizen reports, and sensor data streams — that overwhelm traditional Knowledge Management (KM) frameworks. Conventional AI-driven systems struggle to deliver accurate, traceable responses on complex, multi-domain queries, exhibiting persistent hallucination and insufficient contextual grounding: benchmarking studies reveal that advanced LLMs achieve only 34% accuracy on complex queries, standard Retrieval-Augmented Generation (RAG) improves this to merely 44%, and state-of-the-art solutions answer only 63% of questions without hallucination. These limitations are particularly critical in regulated public governance contexts where transparency, provenance, and temporal consistency are fundamental requirements. This paper presents a Hybrid Retrieval-Augmented Generation (H-RAG) architecture for knowledge management in UDT environments, integrating vector-based semantic retrieval with Knowledge Graph-grounded reasoning. Following a Design Science Research (DSR) methodology, we describe the end-to-end architecture of a Knowledge Management System (KMS) comprising six interconnected modules: heterogeneous document ingestion, parsing and normalisation, transformer-based embedding generation, incremental Knowledge Graph construction, hybrid multimodal indexing, and context-fused retrieval and generation. A Knowledge Intake Guard (KIG) ensures semantic consistency and deduplication prior to graph ingestion, preventing knowledge base fragmentation as the document corpus evolves. The KMS adopts a microservices-oriented pattern with asynchronous processing (message broker and task queue) and multimodal storage (object store, vector database, graph database, and search engine), enabling scalable ingestion of structured and unstructured municipal document corpora. Unlike batch-oriented graph approaches, the proposed system supports incremental knowledge enrichment without requiring periodic full reconstruction of the knowledge base. An illustrative case study based on a municipal traffic ordinance demonstrates how the architecture supports cross-silo knowledge integration, serving both citizens submitting natural-language queries and technical operators requiring entity-centric relational retrieval. Benchmarking evidence from the literature confirms substantial accuracy improvements in Knowledge Graph-augmented retrieval settings, with a corresponding significant reduction in hallucinations. The proposed KMS is developed within the AVANT IPCEI-CIS project and contributes: (i) a conceptual framework for hybrid knowledge retrieval in dynamic urban environments; (ii) a replicable DSR-grounded design approach applicable to public administration KM; and (iii) architectural guidelines and design principles for integrating governance-oriented KM capabilities into multi-silo UDT contexts.