Skip to content

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

Conference: AAAI2026 arXiv: 2508.10419
Code: EternityJune25/ComoRAG
Area: Information Retrieval Keywords: RAG, Long Narrative Reasoning, Memory-Organized, Metacognitive Regulation, Stateful Reasoning

TL;DR

Inspired by the metacognitive regulation mechanism of the prefrontal cortex, this paper proposes the ComoRAG framework, which achieves stateful multi-step reasoning via a dynamic memory workspace and iterative probe queries, significantly outperforming existing RAG methods on long narrative understanding tasks (200K+ tokens).

Background & Motivation

State of the Field

Background: Understanding long narratives (novels, stories) differs from general multi-hop QA in that it requires dynamically constructing and revising a global mental model of plot, characters, and motivational development. Consider the question "Why did Snape kill Dumbledore?"—answering it requires weaving a complete evidence network from scattered clues spanning multiple books: Dumbledore's terminal illness, the Unbreakable Vow, and Snape's deeply concealed loyalty. The true significance of these clues can only be fully reconciled in retrospect.

Shortcomings of existing methods:

Limitations of Prior Work

Limitations of Prior Work: Long-context LLMs are constrained by limited context windows and are prone to the "lost in the middle" problem as input length increases.

Root Cause

Key Challenge: Single-step retrieval RAG methods (e.g., RAPTOR, HippoRAGv2, GraphRAG) rely on static indices, leading to shallow understanding and an inability to capture the dynamic evolution of evidence.

Starting Point

Key Insight: Multi-step retrieval RAG methods (e.g., IRCoT, Self-RAG, MemoRAG) treat each retrieval step independently, lacking coherent narrative reasoning and resulting in fragmented, stateless understanding.

Starting Point

Goal: How can stateful reasoning be realized in a RAG setting—i.e., maintaining a dynamically updated memory throughout the retrieval process so that the system can, like a human reader, continuously construct and revise a global understanding of the narrative as new evidence emerges?

Method

The ComoRAG framework consists of three core components:

1. Hierarchical Knowledge Source

Retrieval indices are established along three complementary cognitive dimensions:

  • Veridical Layer: Retains original text chunks directly and uses an LLM to generate knowledge triples (subject–predicate–object) for each chunk, enhancing query-to-chunk matching precision. This is the most fundamental and important layer.
  • Semantic Layer: Employs RAPTOR's GMM-driven clustering algorithm to recursively aggregate semantically similar chunks into a hierarchical summary tree, capturing thematic and conceptual connections across long distances.
  • Episodic Layer: Produces continuous narrative summaries of text chunks via a sliding window, reconstructing storylines and plot arcs to capture temporal development and causal chains.

2. Dynamic Memory Workspace

A memory unit \(m = (p, \mathcal{E}_p^{type}, \mathcal{C}_p^{type})\) serves as the bridge for multi-step reasoning and contains: - \(p\): the probe query that triggered retrieval - \(\mathcal{E}_p^{type}\): a homogeneous evidence set retrieved from a specific knowledge layer - \(\mathcal{C}_p^{type}\): a synthesized clue generated by the Comprehension Agent, reflecting how the retrieved evidence supplements the understanding of the original query

The memory pool \(\mathcal{M}_{pool}\) is continuously updated at each reasoning cycle, enabling cumulative knowledge accumulation across steps.

3. Metacognitive Regulation Loop

The core cognitive loop comprises six operations:

  1. Self-Probe: When the current round of reasoning fails, the Regulation Agent generates a new probe query \(\mathcal{P}^{(t)}\) based on the original query, the history of probe records, and the synthesized clues from the previous round, opening new exploratory paths.
  2. Tri-Retrieve: For each probe query, dense passage retrieval is performed independently across all three knowledge layers.
  3. Mem-Encode: A memory unit is generated for each retrieval result, recording how new evidence supplements the final answer.
  4. Mem-Fuse: Historically relevant memory clues are retrieved from the memory pool based on semantic similarity to the original query; the Integration Agent then fuses new and existing evidence to produce a high-level contextual summary.
  5. Try-Answer: The QA Agent attempts to answer the original query based on the new evidence from the current round and the fused historical context; if evidence is insufficient, a Failure Signal is issued and the next round begins.
  6. Mem-Update: Newly generated memory units are merged into the global memory pool.

The entire process iterates for at most 5 rounds. The initial step (\(t=0\)) performs one round of Tri-Retrieve + Try-Answer; if it fails, the full metacognitive loop is activated.

Key Experimental Results

Main results on four long-context narrative understanding datasets (all using GPT-4o-mini as the LLM backbone and BGE-M3 as the retrieval model):

Main Results

Method NarrativeQA F1 EN.QA F1 EN.MC ACC DetectiveQA ACC
HippoRAGv2+IRCoT 28.98 29.27 64.19 62.50
RAPTOR+IRCoT 31.35 32.09 63.76 64.77
ComoRAG 31.43 34.52 72.93 68.18

Key findings:

  • On EN.MC, ComoRAG improves accuracy by approximately +8.7% over the strongest baseline, HippoRAGv2+IRCoT.
  • Despite using only the 0.3B BGE-M3 retrieval model, ComoRAG substantially outperforms Naive RAG with an 8B embedding model.
  • On documents exceeding 150K tokens, the accuracy gap between ComoRAG and HippoRAGv2 peaks at +24.6%.
  • Switching to GPT-4.1 raises EN.MC accuracy from 72.93% to 78.17%.
  • When applied as a plug-and-play module on top of RAPTOR, EN.MC accuracy improves from 57.21% to 69.00% (+21% relative gain).
  • Performance gains largely converge within the first 2–3 rounds of iteration.

Ablation Study

Ablation experiments confirm: removing the Veridical Layer causes approximately 30% relative performance degradation; removing the metacognitive process (Metacognition) causes EN.QA F1 to drop by approximately 22%; removing the regulation process (Regulation) causes EN.MC accuracy to drop by approximately 24%.

Highlights & Insights

  • Cognitive Science Inspiration: The framework is designed from the metacognitive regulation mechanism of the human prefrontal cortex, modeling narrative reasoning as a dynamic interplay between "new evidence acquisition" and "integration of prior knowledge," with a theoretically well-motivated design.
  • Stateful Reasoning: Cross-step knowledge accumulation and reasoning state tracking are achieved via the dynamic memory workspace, addressing the fundamental statelessness of traditional multi-step RAG.
  • Three-Layer Knowledge Indexing: Context is modeled across three complementary dimensions—veridical, semantic, and episodic—with the Episodic Layer's design specifically targeting the capture of narrative timelines.
  • Strong Modularity: The core cognitive loop can be applied as a plug-and-play module to existing RAG methods (e.g., RAPTOR, HippoRAGv2) and is compatible with different LLM backbones.
  • Targeted Advantage: On narrative-type queries requiring global context understanding, relative F1 improvement reaches 19%, precisely addressing the bottleneck of existing methods.

Limitations & Future Work

  • Multi-round iteration introduces additional LLM inference overhead (up to 5 rounds, each involving multiple agent calls); actual latency and cost are not sufficiently discussed.
  • Experiments focus primarily on narrative text; generalizability to other long-document types (scientific papers, legal documents, etc.) has not been validated.
  • Construction of the hierarchical knowledge source itself relies on LLM-generated triples and summaries, introducing additional preprocessing overhead.
  • The robustness of the Episodic Layer's sliding window design to varied narrative structures (non-linear narratives, multi-perspective narratives) warrants further investigation.
  • The classification under model compression is inaccurate; the work more appropriately belongs to NLP/RAG-related domains.
Method Knowledge Index Multi-Step Retrieval Memory Mechanism Stateful Reasoning
RAPTOR Semantic Tree ✗ ✗ ✗
HippoRAGv2 Knowledge Graph ✗ ✗ ✗
IRCoT Raw Chunks ✓ (CoT) ✗ ✗
MemoRAG Compressed Context ✓ Dual System Partial
Self-RAG Raw Chunks ✓ (Adaptive) ✗ ✗
ComoRAG Three-Layer Hierarchical ✓ (Iterative Probe) Dynamic Workspace ✓

The core distinction of ComoRAG lies in the introduction of an explicit memory workspace and goal-directed probe query generation, making each retrieval round not independent but a directed exploration grounded in the historical reasoning state.

  • The design philosophy of the dynamic memory workspace is generalizable to other RAG scenarios requiring multi-step reasoning (e.g., code debugging, medical diagnosis).
  • The Episodic Layer design in the three-layer knowledge index offers reference value for tasks such as video understanding and temporal event reasoning.
  • The Self-Probe mechanism in the cognitive loop parallels the "reflection–replanning" paradigm in LLM agents and can be integrated with agent frameworks.
  • The plug-and-play property enables ComoRAG to serve as an enhancement module for existing RAG systems, offering practical engineering value.

Rating

  • Novelty: 8/10 — The cognitively inspired approach to stateful RAG is novel, and the dynamic memory workspace concept demonstrates originality.
  • Experimental Thoroughness: 8/10 — Four datasets, multiple baselines, comprehensive ablations and analyses, though efficiency comparisons are absent.
  • Writing Quality: 8/10 — Analogies are clear, structure is complete, and the Harry Potter example is highly illustrative.
  • Value: 7/10 — The plug-and-play property has engineering value, though the application scope is relatively narrow (narrative understanding).