This August 2025 paper introduces ComoRAG, a novel framework designed to enhance long-context narrative comprehension in Large Language Models (LLMs) by simulating human metacognitive regulation. It addresses the limitations of existing Retrieval-Augmented Generation (RAG) methods, which struggle with stateful reasoning and integrating contradictory evidence over extended narratives. ComoRAG employs a dynamic cognitive loop that includes a hierarchical knowledge source (veridical, semantic, and episodic layers) and a dynamic memory workspace to continuously acquire new evidence and consolidate knowledge. Experimental results demonstrate ComoRAG's superior performance, particularly in solving complex narrative and inferential queries across various datasets, showcasing its robustness and flexibility as a model-agnostic, plug-and-play solution.

Source:

https://arxiv.org/pdf/2508.10419