Memory-Augmented Knowledge Fusion with Safety-Aware Decoding for Domain-Adaptive Question Answering

Domain-specific question answering (QA) systems for service applications face inherent challenges in integrating heterogeneous knowledge sources while ensuring both accuracy and safety. Existing large language models frequently struggle with factual consistency, contextual grounding, and controlled response behavior in sensitive domains such as healthcare policies and government welfare. To address these limitations, we introduce Knowledge-Aware Reasoning and Memory-Augmented Adaptation (KARMA), a novel framework designed to enhance QA performance in real care-driven scenarios. KARMA integrates a dual-encoder architecture to fuse structured and unstructured knowledge, a gated memory unit that dynamically regulates external knowledge injection, and a safety-aware controllable decoder that suppresses unsafe output through classification-driven guidance. Experimental results on a proprietary domain QA dataset show that KARMA significantly outperforms strong baselines in both answer quality and safety robustness. This work presents a comprehensive, adaptable, and safety-aligned solution for building trustworthy QA systems in real-world service contexts.