Paper List

期刊: ArXiv Preprint
发布日期: 2026-03-15
Artificial IntelligenceNatural Language Processing

D-MEM: Dopamine-Gated Agentic Memory via Reward Prediction Error Routing

UC San Diego | Carnegie Mellon University

Yuru Song, Qi Xin
Figure
Figure
Figure
Figure
Figure

30秒速读

IN SHORT: This paper addresses the fundamental scalability bottleneck in LLM agentic memory systems: the O(N²) computational complexity and unbounded API token costs caused by processing every user utterance through expensive memory evolution pipelines, regardless of information value.

核心创新

  • Methodology Introduces D-MEM, a bio-inspired architecture implementing dopamine-gated fast/slow routing based on Agentic Reward Prediction Error (RPE), decoupling short-term interaction from long-term cognitive restructuring.
  • Methodology Develops the LoCoMo-Noise benchmark with systematic 75% noise injection (Filler: 40%, Status: 30%, Tangent: 30%) to simulate real-world conversational dynamics and evaluate memory robustness.
  • Methodology Implements zero-cost retrieval augmentation through hybrid BM25 search with Reciprocal Rank Fusion and an O(1) Shadow Buffer fallback mechanism to prevent adversarial hallucinations.

主要结论

  • D-MEM reduces API token consumption by 80% (from 1,648K to 319K tokens) while maintaining or improving accuracy on complex reasoning tasks under extreme noise conditions (ρ=0.75).
  • The architecture achieves superior multi-hop reasoning performance (42.7% F1 vs. A-MEM's 27.0%, a +15.7 point gap) by preserving cleaner relational memory structures through selective cognitive restructuring.
  • The Critic Router successfully gates 80% of computational resources while maintaining overall F1 score of 37.4% on standard benchmarks, demonstrating the effectiveness of the bio-inspired RPE mechanism.
研究空白： Current evolving memory frameworks (e.g., A-MEM) operate as synchronous "append-and-evolve-all" systems, indiscriminately processing all inputs through O(N²) pipelines, leading to severe write-latency, unbounded API costs, and context pollution from conversational noise.

摘要: The integration of structured, long-term memory is critical for the development of autonomous Large Language Model (LLM) agents. Recent advancements, such as the Agentic Memory (A-MEM) framework, have achieved significant progress by dynamically constructing and evolving knowledge graphs. However, existing architectures inherently operate as synchronous, "append-and-evolve-all" systems. Processing every user utterance through a computationally expensive O(N²) memory evolution pipeline introduces severe write-latency, unbounded API token costs, and catastrophic context window pollution caused by conversational noise. To address this scalability bottleneck, we introduce D-MEM (Dopamine-Gated Agentic Memory), a biologically inspired architecture that decouples short-term interaction from long-term cognitive restructuring. Drawing inspiration from the Dopamine-driven Reward Prediction Error (RPE) gating mechanism in the mammalian Ventral Tegmental Area (VTA), D-MEM implements a highly efficient Fast/Slow routing system. We introduce a lightweight Critic Router that continuously evaluates the Information Entropy (Surprise) and Long-term Utility of incoming stimuli. Routine inputs with low RPE are either bypassed entirely or cached in an O(1) fast-access buffer, preserving computational resources. Conversely, inputs generating a high RPE—such as factual contradictions or paradigm-shifting preference changes—trigger a "dopamine release" that activates the slow, O(N) deep memory evolution pipeline, actively reshaping the agent's global knowledge graph. To enable rigorous evaluation under realistic conditions, we further introduce the LoCoMo-Noise benchmark, which systematically injects controlled conversational noise into long-term dialogue sessions to simulate real-world interaction dynamics. Extensive evaluations demonstrate that D-MEM reduces API token consumption by over 80% and eliminates O(N²) write-latency bottlenecks, all while strictly outperforming synchronous baselines in complex multi-hop reasoning and adversarial resilience. By selectively gating cognitive restructuring and leveraging zero-cost retrieval augmentations, D-MEM provides a highly scalable and cost-efficient foundation for lifelong agentic memory. To support reproducibility, we open-source our implementation at https://github.com/london-and-tequila/dmem.

代码