ComoRAG Framework: Cognitive Memory RAG

• ComoRAG is a cognitive-inspired, stateful RAG framework that uses iterative cycles of memory consolidation and evidence retrieval to perform advanced narrative reasoning.
• It leverages a three-layer hierarchical knowledge source and dynamic memory workspace to integrate factual, thematic, and episodic cues from ultra-long narrative texts.
• Empirical evaluations demonstrate up to 11% relative improvements on long narrative benchmarks, highlighting its effectiveness in global and multi-chapter reasoning tasks.

ComoRAG is a cognitive-inspired, memory-organized Retrieval-Augmented Generation (RAG) framework designed for stateful long narrative reasoning, particularly over extended storylines and novels. Traditional RAG approaches, which perform stateless single-pass retrieval over external corpora, often fail to resolve complex queries that require global comprehension across hundreds of thousands of tokens. ComoRAG addresses this challenge by modeling narrative reasoning as an iterative, dynamically evolving cycle between evidence acquisition and memory consolidation, drawing direct inspiration from human metacognitive regulation.

1. Framework Architecture and Design Principles

ComoRAG departs from conventional RAG by introducing a stateful, cycle-based architecture for retrieval and reasoning. Each reasoning session consists of multiple iterative cycles, during which the system dynamically interacts with a memory workspace and hierarchical knowledge sources. The framework’s architecture comprises three foundational components:

• Hierarchical Knowledge Source: Raw narrative texts are partitioned into three semantic layers:
  • Veridical Layer: Stores factual, chunk-level evidence, often in the form of knowledge triples.
  • Semantic Layer: Captures thematic abstractions by clustering and summarizing topical segments.
  • Episodic Layer: Reconstructs narrative flow through sequential or window-based summarization.
• Dynamic Memory Workspace: Acts as a repository for “memory units” $(m)$, each formally expressed as $$m = (p, \mathcal{A}^{(\mathrm{type})}_p, \mathcal{C}^{(\mathrm{type})}_p)$$, where $p$ is a probing query, $\mathcal{A}^{(\mathrm{type})}_p$ the retrieved evidence, and $\mathcal{C}^{(\mathrm{type})}_p$ a synthesized cue for answer support.
• Metacognitive Control Loop: Orchestrates iterative reasoning cycles, alternating between generating targeted probing queries and consolidating newly retrieved evidence into the global memory pool.

These mechanisms collectively enable the system to revisit knowledge, refine understanding, and build coherent mental models akin to human narrative comprehension.

2. Cognitive Inspiration and Metacognitive Regulation

ComoRAG’s design is grounded in cognitive neuroscience, specifically modeling the regulatory and memory-related functions of the prefrontal cortex (PFC). Human narrative reasoning is not performed in a single step but unfolds as an adaptive sequence of evidence gathering, integration, and reflective self-probing.

• Regulatory Process: When encountering a reasoning impasse, ComoRAG generates new probing queries via a self-reflective strategy:

$$\mathcal{P}^{(t)} = \pi_{\mathrm{probe}}(q_{\mathrm{init}}, \mathcal{P}_{\mathrm{hist}}^{(t-1)}, \{\mathcal{C}\}^{(t-1)})$$

where $q_{\mathrm{init}}$ is the original question, $\mathcal{P}_{\mathrm{hist}}^{(t-1)}$ is the history of previous probes, and $\{\mathcal{C}\}^{(t-1)}$ are consolidated cues.

• Iterative Memory Consolidation: Newly encoded memory units are incrementally merged into the workspace:

$$\mathcal{M}^{(t)}_{\mathrm{pool}} \leftarrow \mathcal{M}^{(t-1)}_{\mathrm{pool}} \cup \mathcal{M}^{(t)}_{\mathrm{encode}}$$

promoting a stateful integration of context and supporting dynamic, non-static retrieval logic.

This paradigm mirrors human mental modeling, where evidence is gathered and reinterpreted as narrative comprehension progresses.

3. Iterative Reasoning Cycle and Stateful Retrieval

ComoRAG operationalizes stateful reasoning through tightly coupled retrieval and memory update cycles:

1. Initial Retrieval: Attempts to answer the query with conventional one-shot retrieval over the narrative corpus.
2. Regulatory Reasoning: If resolution is unsatisfactory, the agent invokes $\pi_{\mathrm{probe}}$ to generate alternative queries that explore knowledge gaps.
3. Layered Retrieval: For each probe, evidence is retrieved from all hierarchical layers, ensuring coverage of factual, thematic, and episodic cues.
4. Memory Encoding: Retrieved evidence and corresponding probes are structured into memory units.
5. Memory Fusion: Similar or relevant past memory units are merged using a fusion operator:

$$\mathcal{C}_{\mathrm{fuse}}^{(t)} = \pi_{\mathrm{fuse}}(q_{\mathrm{init}}, \mathcal{M}^{(t-1)}_{\mathrm{pool}} \circ q_{\mathrm{init}})$$

integrating historical context relevant to the original query.

1. Answer Generation: The QA agent $\pi_{\mathrm{QA}}$ leverages both newly retrieved evidence and fused memory to attempt a confident answer.
2. Termination Criterion: Cycles continue (typically 2–3 iterations) until resolution or reaching a maximum iteration cap.

This structured loop grants ComoRAG the ability to perform deep inferential reasoning requiring cross-chapter and cross-entity integration.

4. Dynamic Memory Workspace

Central to ComoRAG’s narrative reasoning is a continuously updated, global memory workspace:

• Memory Unit Formalization: Each memory unit stores the tuple $$(p, \mathcal{A}^{(\mathrm{type})}_p, \mathcal{C}^{(\mathrm{type})}_p)$$, where $p$ is a probing query, $\mathcal{A}$ the retrieved evidence from a given layer, and $\mathcal{C}$ a synthesized cue aligning the evidence to the global answer intent.
• Memory Pool Operations: At each cycle, new units are encoded via $\mathrm{memcolor}$ and merged via set union (see formula above), allowing persistent accumulation and reuse of relevant evidence and cues.

This evolving workspace preserves reasoning states and acts as a repository for the agent’s “mental model,” directly analogous to short- and long-term human memory during reading.

5. Performance on Long-Narrative Benchmarks

Empirical evaluation was conducted on four long-context narrative benchmarks (NarrativeQA, EN.QA, EN.MC, DetectiveQA), with context lengths exceeding 200K tokens.

• Relative Gains: ComoRAG delivers consistent relative improvements, up to 11% over the strongest baseline for long narrative comprehension tasks.
• Accuracy: For example, on EN.MC, performance increased from ~64–65% (baseline frameworks) to 72.93%.
• Domain Advantages: The stateful reasoning mechanism particularly benefits “global queries” requiring full-narrative integration. F1 score improvements in narrative/inferential queries range from 16–19%.

These results underscore the framework’s efficacy when extensive plotlines, evolving relations, or multi-chapter entity dynamics are involved.

6. Applications and Implications

ComoRAG is extensible to a diverse array of tasks that require global, stateful narrative reasoning.

• Literary Analysis: Enables comprehensive plot and character analysis across entire novels.
• Complex Question Answering: Supports resolution of inferential and multi-dimensional queries over extended documents.
• Educational Tech: Facilitates systems that assist in literature comprehension by modeling story evolution and summary extraction.
• Legal/Technical Review: A plausible implication is application to stateful reasoning over lengthy legal or technical documents, given its ability to continually refine evidence pools and revisit context.

The cognitive-inspired, memory-organized paradigm suggests that future retrieval frameworks may increasingly structure reasoning as an interplay of evidence accumulation and memory integration.

7. Code Availability and Implementation

Implementation is publicly released: [https://github.com/EternityJune25/ComoRAG]. Technical details outlined in the original work include:

• Hierarchical Source Construction: Detailed instructions for knowledge graph extraction, window-based summarization, and semantic clustering.
• Iterative Reasoning Loop: Provided pseudocode formalizes each operator (probe generation, retrieval, encoding, fusion, answer attempt).
• Hyperparameters: Full specification for context length, cycle limits, and memory allocation ratios.
• Plug-and-Play Flexibility: The metacognitive loop can be integrated into other retrieval frameworks (e.g., RAPTOR, HippoRAGv2), enhancing their performance.

This open-source release enables direct reproduction and extension for research and practical applications.

---

ComoRAG represents an advanced, cognitively-motivated framework for stateful, memory-driven narrative reasoning in retrieval-augmented generation systems. By operationalizing reasoning as an iterative dialog between memory and evidence acquisition, it substantially improves global comprehension and complex query resolution over ultra-long contexts, and establishes a blueprint for future RAG paradigms that more closely mirror human cognition (Wang et al., 14 Aug 2025).