This study presents MedRACE-L3, a hybrid framework that integrates Large Language Models (LLMs) within a Retrieval-Augmented Generation (RAG) paradigm, complemented by Maximal Marginal Relevance (MMR) and Agentic Context Engineering (ACE), to generate concise, accurate and clinically meaningful summaries of patient medical records. Standalone LLMs often exhibit limitations, including incomplete contextual understanding, factual inconsistencies and redundant content. RAG dynamically incorporates relevant clinical documents and external knowledge sources into the generation process, while ensuring traceability by explicitly linking generated outputs to their originating evidence, thereby improving factual consistency, domain alignment and transparency. Retrieved reports are then evaluated and selected using MMR, which promotes diversity and reduces redundancy. ACE dynamically scores query–example relevance to guide context selection, using high-scoring cases for in-context learning and playbook constraints for lower-scoring ones to improve generation quality. Evaluation relies on standard automatic metrics, with BLEU and ROUGE measuring surface-level similarity and BERTScore assessing semantic similarity between generated summaries and ground-truth references. Experimental results demonstrate that MedRACE-L3 significantly improves accuracy, completeness, coherence, interpretability and trustworthiness compared to baseline models, highlighting the effectiveness of combining context optimization, diversity-aware retrieval, score-driven contextual adaptation and traceable generation for reliable and interpretable clinical decision-support systems.