This thesis explores the integration of knowledge into Task-Oriented Dialogue (TOD) systems within Intelligent Assistants (IA) to enhance their ability to understand user intent and execute structured tasks. Despite advancements in Natural Language Processing (NLP) and Large Language Models (LLMs), TOD systems continue to face challenges related to knowledge integration, scalability, and generalization. This research proposes methodologies for incorporating domain-specific, dialogue-level, and cross-lingual knowledge into TOD systems, thereby improving their adaptability and effectiveness across diverse applications. To address key challenges in knowledge integration, this study introduces innovative methods such as structured slot-value transfer and schema-guided knowledge graphs to incorporate domain-specific knowledge for Dialogue State Tracking (DST). Furthermore, the research explores dialogue-level knowledge integration to improve context awareness by incorporating dialogue acts into slot-value prediction and employing a soft mixture-of-experts approach. These techniques enhance comprehension, improve efficiency, and contribute to the scalable development of TOD systems. Additionally, the thesis presents a cross-lingual knowledge transfer mechanism to improve commonsense reasoning in low-resource languages, enhancing the multilingual adaptability of TOD systems. In summary, this work systematically incorporates knowledge at different levels and evaluates its impact across diverse contexts to enhance the design of more effective and intelligent conversational agents. By equipping TOD systems with dynamic knowledge integration, this research not only advances the development of more robust, scalable, and adaptable conversational agents but also improves user interaction, enhances response accuracy, and broadens the applicability of TOD systems across various domains and languages.