RippleMem: From Isolated Retrieval to Associative Recollection for Long-Term Agent Memory
RippleMem uses event-centric memory graphs and associative recall, improving accuracy by 11.87% and reducing graph construction costs by 30x.
Key Findings
Methodology
RippleMem constructs an event-centric memory graph, transforming interaction history into cue-rich episodic units linked by semantic and structural associations. During writing, dialogue windows are processed into structured memory units with contextual cues (time, location, participants), stored in a sparse graph. During reading, multi-modal cues (semantic, lexical, structural) initiate initial retrieval, followed by local expansion along association paths to recover missing supporting evidence. This process mimics human cue-dependent recall, enabling evidence completion rather than one-shot retrieval, thus addressing noise and incomplete record issues prevalent in prior methods.
Key Results
- On LoCoMo, RippleMem improved LLM-as-Judge accuracy by 3.95%, reaching 52.49%. On LongMemEval-S, the improvement was up to 11.87%. It also reduced graph construction costs by approximately 30 times compared to graph-based baselines, demonstrating high efficiency and accuracy.
- Ablation studies confirmed that event-centric graph structure and multi-modal associative recall are key to performance gains. The multi-path expansion strategy significantly enhances long-distance evidence recovery, especially in multi-turn, complex scenarios.
- Across different task complexities, RippleMem showed robust performance and scalability, outperforming existing methods in multi-task and multi-user environments, indicating strong generalization and practical applicability.
Significance
This work advances long-term memory retrieval by integrating cognitive principles of cue-dependent and event-based organization, significantly improving multi-step reasoning and evidence integration. Its innovative memory graph and associative recall mechanisms provide a foundation for more cognitively plausible AI systems, addressing longstanding challenges in scalable, accurate, and cost-effective memory management for complex AI applications.
Technical Contribution
The paper introduces a novel event-centric memory graph combined with multi-modal association pathways, enabling adaptive evidence recovery. It innovates by modeling evidence-conditioned recall, inspired by human cognition, and optimizing retrieval paths to balance efficiency and completeness. This approach surpasses prior flat or purely structural methods, offering a scalable, high-precision memory system with theoretical and engineering advantages.
Novelty
This is the first integration of event-based memory graphs with associative, evidence-conditioned recall, bridging cognitive science insights with scalable AI memory architectures. Unlike prior work that relies on static or flat retrieval, RippleMem dynamically expands support paths based on cues, closely mimicking human layered recall, representing a significant leap in memory system design.
Limitations
- In scenarios with extremely sparse or noisy memory cues, the system may struggle to recover complete evidence, especially if key associations are missing or poorly grounded.
- While the sparse graph construction reduces costs, in ultra-large-scale environments, computational overhead for neighbor expansion and scoring remains non-trivial, requiring further optimization.
- Current reliance on structured cues and semantic similarity may limit performance in multi-modal or unstructured data environments, necessitating future integration of richer data modalities and adaptive updating mechanisms.
Future Work
Future efforts will focus on integrating multi-modal data (visual, auditory) to enrich episodic units, enabling cross-modal associative recall. Developing dynamic memory updating strategies and reinforcement learning-based control for recall paths will improve robustness. Expanding to multi-task, multi-user settings and real-world deployment scenarios will further validate and enhance system capabilities.
AI Executive Summary
Long-term memory management remains a fundamental challenge in deploying large language models for complex reasoning tasks. Existing retrieval methods, such as full-context search or flat retrieval, often suffer from noise, incomplete evidence, or high computational costs, limiting their effectiveness in multi-turn, long-horizon interactions. RippleMem addresses these issues by introducing an event-centric memory graph that organizes interaction history into cue-rich, structured units. During inference, it employs a multi-modal associative recall process, starting from initial retrieval cues and expanding along semantic and structural associations to recover missing evidence. This evidence-conditioned approach closely mimics human memory, where recall is driven by cues and associations rather than direct matching alone. Experimental results on LoCoMo and LongMemEval-S datasets demonstrate that RippleMem achieves the highest overall performance, with improvements of 3.95% and 11.87% in accuracy, respectively, compared to strong baselines. Additionally, it reduces graph construction costs by approximately 30 times, showcasing its efficiency. These advances significantly enhance the capability of AI agents to perform multi-step reasoning, evidence integration, and long-term interaction management. The system's design aligns with cognitive theories of episodic memory and associative recall, providing a scalable, interpretable, and effective framework for future long-term memory systems. Moving forward, integrating multi-modal data, dynamic memory updating, and multi-user support will broaden RippleMem’s applicability, pushing AI closer to human-like long-term cognition and reasoning.
Deep Dive
Abstract
LLM-based agents increasingly rely on external memory to support long-horizon reasoning and interaction. However, the main bottleneck is not simply storing past experience, but recovering the right set of evidence when relevant information is distributed across many interactions. Existing approaches struggle with this access problem. Full-context methods require noisy long-context search, flat retrieval often returns isolated and incomplete records, and graph-based memory systems can be expensive to construct while compressing rich event context. We introduce RippleMem, a long-term memory system that replaces one-shot retrieval with adaptive associative recollection. Inspired by cue-dependent episodic retrieval and associative completion, RippleMem stores interaction history as cue-rich episodic memory units and organizes them in an event-centric memory graph. Given a query, it first recalls relevant memory anchors through hybrid cues, then expands from these anchors along semantic and structural associations to recover missing supporting evidence. In this way, initially recalled memories serve not only as answer context, but also as cues for completing the evidence needed to answer. Experiments on LoCoMo and LongMemEval-S show that RippleMem achieves the best overall performance across evaluated settings, improving LLM-as-a-Judge accuracy by 3.95% on LoCoMo and up to 11.87% on LongMemEval-S, while reducing graph construction cost by about 30x.