ComoRAG: A Cognitive-Inspired Memory-Organized RAG for Stateful Long Narrative Reasoning

TL;DR

ComoRAG addresses the difficulty of long-narrative understanding by moving beyond stateless, one-shot retrieval. It introduces a cognitive-inspired framework with a memory-organized dynamic loop that actively probes, retrieves, and integrates evidence from a hierarchical knowledge source (veridical, semantic, episodic layers) within a dynamic memory workspace. Through metacognitive regulation, the approach achieves stateful reasoning over long contexts and demonstrates consistent gains over strong baselines across four challenging datasets, with improvements most pronounced on complex, globally-contextual queries. The results indicate strong modularity and generalization: the metacognitive loop can enhance existing RAG systems and benefit from stronger LLM backbones, providing a scalable paradigm for retrieval-based long-context narrative reasoning.

Abstract

Narrative comprehension on long stories and novels has been a challenging domain attributed to their intricate plotlines and entangled, often evolving relations among characters and entities. Given the LLM's diminished reasoning over extended context and its high computational cost, retrieval-based approaches remain a pivotal role in practice. However, traditional RAG methods could fall short due to their stateless, single-step retrieval process, which often overlooks the dynamic nature of capturing interconnected relations within long-range context. In this work, we propose ComoRAG, holding the principle that narrative reasoning is not a one-shot process, but a dynamic, evolving interplay between new evidence acquisition and past knowledge consolidation, analogous to human cognition on reasoning with memory-related signals in the brain. Specifically, when encountering a reasoning impasse, ComoRAG undergoes iterative reasoning cycles while interacting with a dynamic memory workspace. In each cycle, it generates probing queries to devise new exploratory paths, then integrates the retrieved evidence of new aspects into a global memory pool, thereby supporting the emergence of a coherent context for the query resolution. Across four challenging long-context narrative benchmarks (200K+ tokens), ComoRAG outperforms strong RAG baselines with consistent relative gains up to 11% compared to the strongest baseline. Further analysis reveals that ComoRAG is particularly advantageous for complex queries requiring global context comprehension, offering a principled, cognitively motivated paradigm towards retrieval-based stateful reasoning. Our framework is made publicly available at https://github.com/EternityJune25/ComoRAG.

Figures (6)

• Figure 1: Comparison of RAG reasoning paradigms.
• Figure 2: An illustration of ComoRAG. Triggered by a reasoning impasse (Failure), the Metacognitive Regulation loop consists of five core operations described in Section \ref{['ssec:loop']}: 1) [draw=selfprobecolor, thick, inner sep=2.2pt]Self-Probe Self-Probeto devise new exploratory probing queries based on past memory units; 2) [draw=black, thick, inner sep=2.2pt]Tri-Retrieve Tri-Retrieveto retrieve evidence from three knowledge sources; 3) [draw=memcolor, thick, inner sep=2.2pt]Mem-Encode Mem-Encodeto form new memory units on how the latest evidence of new aspects could complement the final query resolution; 4) [draw=memfusecolor, thick, inner sep=2.2pt]Mem-Fuse Mem-Fuseto generate cues integrating new and past memory units; 5) [draw=tryanswercolor, thick, inner sep=2.2pt]Try-Answer Try-Answerto perform query answering using new memory information produced in this cycle.
• Figure 3: Averaged accuracy across different document lengths on Multi-Choice datasets. ComoRAG is shown more robust to long contexts over the baseline.
• Figure 4: Performance gains from iterative probing. GPT-4.1 marks the evaluation by using the stronger GPT-4.1 as LLM agents in ComoRAG (as opposed to GPT-4o-mini).
• Figure 5: Distribution of solved query types.