LiCoMemory: Lightweight and Cognitive Agentic Memory for Efficient Long-Term Reasoning
LiCoMemory uses hierarchical CogniGraph for lightweight, structured long-term memory, boosting reasoning efficiency by 23%.
Key Findings
Methodology
This paper introduces LiCoMemory, an end-to-end framework integrating CogniGraph—a lightweight hierarchical graph that decouples semantic indexing from storage. It employs entity-relation triples as semantic anchors, combined with temporal and hierarchical-aware retrieval mechanisms, enabling real-time updates and structured knowledge access. The system performs top-down retrieval guided by session summaries, triple relevance, and temporal decay, with integrated reranking to ensure coherence. Experiments on LoCoMo and LongMemEval benchmarks demonstrate significant improvements in accuracy (up to 23%) and retrieval speed, outperforming baselines like GraphRAG and Mem0. The architecture supports incremental graph updates, maintaining a compact, dynamic knowledge structure suitable for multi-session long-term reasoning.
Key Results
- On LongMemEval, LiCoMemory achieved 69.2% accuracy and 72.4% recall, surpassing the second-best by over 20%, with query latency reduced to 1.62 seconds, demonstrating both high effectiveness and efficiency.
- In LoCoMo, it outperformed baselines with a 64.5% accuracy in multi-session temporal reasoning, maintaining low token consumption and response latency, validating its scalability.
- Ablation studies confirmed that hierarchical retrieval and temporal-aware reranking are critical, as removing these components led to performance drops, highlighting their importance.
Significance
This work addresses the inefficiencies of flat or entangled graph-based memory systems by proposing a layered, semantic index approach. It effectively reduces redundancy, improves retrieval relevance, and enables real-time updates, thus advancing long-term conversational reasoning. Its impact extends to intelligent assistants, knowledge management, and multi-turn dialogue systems, offering a scalable solution for persistent, coherent, and contextually grounded AI interactions.
Technical Contribution
Key innovations include the design of CogniGraph as a decoupled, hierarchical semantic index, combined with a temporal and hierarchical-aware retrieval mechanism. The system supports incremental graph updates, fast reranking, and structured knowledge access, providing a new paradigm for dynamic long-term memory. Compared to existing flat or entangled graph models, LiCoMemory offers superior efficiency, interpretability, and adaptability, with theoretical guarantees on retrieval relevance and update speed.
Novelty
This is the first work to embed a hierarchical, semantic-aware graph as a lightweight index layer for long-term memory in conversational agents. Unlike prior flat or entangled graph models, it decouples storage from semantic indexing, enabling fast, structured, and adaptive knowledge retrieval, thus opening new avenues for scalable, real-time long-term reasoning.
Limitations
- The reliance on entity-relation extraction may introduce errors in complex or noisy dialogues, affecting retrieval accuracy.
- Scalability to extremely large knowledge bases or highly dynamic environments remains to be validated, requiring further optimization.
- Real-time performance under high concurrency needs enhancement, especially for very long sessions or rapid updates.
Future Work
Future research will explore multi-modal integration, adaptive graph restructuring, and reinforcement learning-based retrieval optimization. Enhancing robustness in noisy or large-scale scenarios, and extending the framework to multi-agent systems, are also promising directions.
AI Executive Summary
Recent advances in large language models have unlocked impressive conversational and reasoning capabilities. However, their limited context window and lack of persistent memory hinder long-term reasoning, especially in multi-session dialogues. Traditional external memory architectures, often based on flat or entangled graph structures, suffer from redundancy, slow updates, and unstructured retrieval, constraining their practical deployment.
This paper introduces LiCoMemory, a novel framework designed to overcome these limitations. Central to this system is CogniGraph, a lightweight hierarchical graph that acts as a semantic index rather than a static repository. By decoupling knowledge storage from semantic organization, CogniGraph enables efficient, incremental updates and structured retrieval. The architecture employs entity-relation triples as semantic anchors, combined with temporal and hierarchical-aware search and reranking mechanisms, ensuring relevant, coherent knowledge is retrieved in real-time.
Experimental results on benchmarks like LoCoMo and LongMemEval demonstrate that LiCoMemory outperforms existing methods, achieving up to 23% higher accuracy and significantly lower latency. Its ability to maintain multi-session consistency and perform accurate temporal reasoning marks a substantial step forward for long-term conversational AI. The system’s lightweight design allows fast updates, making it suitable for real-world, dynamic environments.
Overall, LiCoMemory offers a scalable, efficient solution for persistent memory in large language models, with broad implications for intelligent assistants, knowledge management, and beyond. Future work will focus on multi-modal integration, adaptive graph restructuring, and reinforcement learning to further enhance its robustness and scalability, paving the way for truly long-term, context-aware AI systems.
Deep Dive
Abstract
Large Language Model (LLM) agents exhibit remarkable conversational and reasoning capabilities but remain constrained by limited context windows and the lack of persistent memory. Recent efforts address these limitations via external memory architectures, often employing graph-based representations, yet most adopt flat, entangled structures that intertwine semantics with topology, leading to redundant representations, unstructured retrieval, and degraded efficiency and accuracy. To resolve these issues, we propose LiCoMemory, an end-to-end agentic memory framework for real-time updating and retrieval, which introduces CogniGraph, a lightweight hierarchical graph that utilizes entities and relations as semantic indexing layers, and employs temporal and hierarchy-aware search with integrated reranking for adaptive and coherent knowledge retrieval. Experiments on long-term dialogue benchmarks, LoCoMo and LongMemEval, show that LiCoMemory not only outperforms established baselines in temporal reasoning, multi-session consistency, and retrieval efficiency, but also notably reduces update latency. Our official code and data are available at https://github.com/EverM0re/LiCoMemory.