ComoRAG: Memory-Augmented RAG for Narratives

• ComoRAG is a cognitive-inspired memory-augmented RAG framework that employs iterative retrieval and memory fusion to improve long narrative comprehension.
• It uses a dynamic memory workspace with probing, hierarchical retrieval, and cue-based fusion to integrate evidence and overcome stateless limitations.
• Empirical evaluations show up to a 24.6% accuracy gain on ultra-long narratives, highlighting its effectiveness in complex, stateful reasoning tasks.

ComoRAG is a cognitive-inspired, memory-organized retrieval-augmented generation (RAG) framework developed for stateful reasoning over long narratives. It addresses the inherent limitations of stateless, single-step retrieval architectures in complex, extended-document reasoning tasks—where evolving, intertwined relationships among entities and events must be modeled for high-fidelity comprehension. ComoRAG draws directly on models of human metacognition, specifically the iterative interplay between memory consolidation and exploratory evidence gathering posited in the neuroscience of the Prefrontal Cortex (PFC). Its innovation lies in orchestrating a memory workspace and iterative metacognitive control loop for dynamic, evolving context construction and resolution, by which it consistently outperforms conventional RAG systems on long-context narrative comprehension (Wang et al., 14 Aug 2025).

1. Theoretical Model and Cognitive Foundations

ComoRAG is grounded in the hypothesis that narrative comprehension requires a process akin to metacognitive regulation—where sequential interactions between new evidence gathering (exploration) and memory consolidation (integration) advance the mental model toward coherence. This theoretical stance is instantiated by departing from stateless, single-step retrieval and instead operationalizing reasoning as an evolving loop: upon encountering a “reasoning impasse,” the system triggers a cycle of “probing,” “retrieval,” “memory fusion,” and “answer refinement.” This architecture is inspired by neuroscientific principles of PFC-driven metacognitive cycles, where memory-related neural signals modulate the acquisition and reorganization of relevant knowledge in the face of uncertainty.

2. Dynamic Memory Workspace and State Representation

Central to ComoRAG is the design of a dynamic memory workspace. Memory is structured as a pool $\mathcal{M}_{\text{pool}}$ containing memory units $$m = (p, \mathcal{E}^{\text{type}}_{p}, \mathcal{C}^{\text{type}}_{p})$$:

• $p$: the probing query responsible for the retrieval event
• $\mathcal{E}^{\text{type}}_{p}$: retrieved evidence, classified according to knowledge granularity—veridical (factual span), semantic (abstracted event/summary), or episodic (narrative flow)
• $\mathcal{C}^{\text{type}}_{p}$: a learned cue encoding how this evidence narrows the uncertainty relative to the original query.

The workspace persists across cycles, enabling the agent to not merely accrete evidence but to maintain a causally structured, ever-evolving context state.

3. Iterative Metacognitive Control Loop

When initial retrievals are inadequate, ComoRAG enters a structured meta-reasoning loop composed of:

1. Probe Generation (selfprobecolor): Synthesize new exploratory queries $$\mathcal{P}^{(t)} = \pi_{\text{probe}}(q_{\text{init}}, \mathcal{P}_{\text{hist}}^{(t-1)}, \{\mathcal{C}\}^{(t-1)})$$ using the current question, historical probes, and cues.
2. Hierarchical Retrieval (black): Retrieve evidence across multi-level knowledge sources for each new probe, targeting both granular factoids and higher-order events/summaries.
3. Memory Unit Construction (memcolor): Build new memory units from retrieved evidence.
4. Memory Fusion (memfusecolor): Integrate new evidence with the existing memory pool, generating a fused cue:

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

1. Answer Attempt (tryanswercolor): Generate an answer $$O^{(t)} = \pi_{QA}(q_{\text{init}}, \mathcal{M}_{\text{encode}}^{(t)}, \mathcal{C}_{\text{fuse}}^{(t)})$$ conditioned on all accumulated evidence and cues.
2. Memory Pool Update:

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

The iteration halts when the synthesized answer satisfies a stopping criterion (e.g., confidence, coverage, or explicit user feedback).

This design operationalizes an active inference process: when context is insufficient, ComoRAG explores new evidence paths adaptively, integrating them in light of prior memory to progressively approach a coherent resolution.

4. Empirical Evaluation and Performance

ComoRAG has been tested across four long-context narrative benchmarks (each exceeding 200K tokens per document): NarrativeQA, EN.QA, EN.MC, and DetectiveQA. It consistently outperforms strong RAG baselines, with documented improvements:

• On EN.MC, accuracy rises from 64.6% (static retrieval) to 72.93% (ComoRAG), an 11% relative gain.
• For complex global comprehension tasks (requiring thread-level or multi-entity narrative integration), F1 improvement reaches 19% over baselines.
• For ultra-long contexts (>150K tokens), ComoRAG’s advantage peaks at +24.6% accuracy.

These gains are attributed directly to ComoRAG’s ability to maintain and update a globally coherent representation of the narrative, counteracting context fragmentation and retrieval myopia present in standard architectures.

5. Applications and Generalization

ComoRAG is tailored to settings where long-range narrative understanding is paramount, including:

• Literary and narrative analysis (comprehension of novels, detective stories, long-form arguments).
• Multi-hop question answering over extended documents.
• Domains requiring consistent state tracking and multi-entity modeling (legal, historical, or biomedical texts).

Its memory-centric design allows for principled navigation of highly entangled, distributed evidence—an essential capacity in cases where single-shot or shallow retrieval falls short.

6. Limitations and Research Directions

The ComoRAG framework, while outperforming prior RAG models, introduces novel complexity and new research questions:

• The construction and synthesis of probing queries and cues—$\pi_{\text{probe}}, \pi_{\text{fuse}}$—may require further tuning for generalization to non-narrative or multi-modal domains.
• Memory pool growth and redundancy management could present practical challenges as iterated reasoning cycles accumulate evidence.
• Potential integration with larger, more capable LLM agents (e.g., GPT-4.1, Qwen3-32B) and more sophisticated memory consolidation or cue abstraction remain active areas for extension.

Future work is suggested in extending ComoRAG to embrace richer modality integration, cross-document or multi-source fusion, and adaptive stopping criteria as well as exploring more biologically plausible memory models for iterative reasoning.

7. Significance and Impact

ComoRAG advances the state of retrieval-augmented generation by encoding the insight that reasoning—especially in narrative domains—is fundamentally a stateful, incremental process, not a passively static or context-agnostic operation. Its cognitive-inspired memory workspace and iterative metacognitive control offer a significant methodological shift in RAG research, showing that “stateful” architectures can yield notably stronger performance on tasks requiring global comprehension, integration across long and complex input spaces, and dynamic, multi-step inference (Wang et al., 14 Aug 2025).