Existing large-scale language models (LLMs) based on Retrieval-Augmented Generation (RAG) are limited to static single-turn interactions and a fixed set of tools, making them unsuitable for dynamic environments such as healthcare and smart homes. In this paper, we present Dynamic Context Tuning (DCT), a lightweight framework that extends RAG to support multi-turn conversations and changing tool environments without retraining. DCT integrates an attention-based context cache to track relevant historical information, LoRA-based retrieval to dynamically select domain-specific tools, and efficient context compression to keep inputs within LLM context constraints. Synthetic and real-world benchmark experiments show that DCT improves planning accuracy by 14% and reduces hallucinations by 37%, while achieving GPT-4 performance at a much lower cost. Furthermore, DCT generalizes to previously unseen tools, enabling scalable and adaptable AI assistants in diverse dynamic environments.
