H-SEAM: A Hierarchical Self-Evolving Agentic Memory System

ACL ARR 2026 January Submission4740 Authors

Keywords: Large Language Models, Long-term Memory, Retrieval-Augmented Generation, AI Agents, Graph-based Memory, Hierarchical Indexing, Dialogue Systems

Abstract: Equipping LLM-based agents with evolving long-term memory remains a persistent challenge. Existing approaches predominantly rely on flat vectors or graphs with predetermined schemas, inevitably leading to semantic fragmentation or structural rigidity. To address this, we propose H-SEAM, a dynamic multi-level memory framework grounded in Dynamic Memory Evolution Theory. Unlike static hierarchies, H-SEAM features: 1) a multi-granular hierarchical structure that vertically integrates raw episodic fragments with semantic abstractions to bridge semantic gaps; 2) a density-driven evolutionary mechanism that dynamically distills high-level abstractions from information saturation, allowing the memory topology to adaptively grow with user interactions; and 3) a sufficiency-guided 2-hop expansion strategy offering the optimal accuracy-efficiency trade-off. Experiments on LOCOMO and LongMemEval demonstrate that H-SEAM consistently surpasses strong baselines (e.g., MemOS), improving overall accuracy by 6.7% while maintaining stable efficiency.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: LLM/AI agents, retrieval-augmented generation, conversational modeling, graph-based methods, knowledge graphs

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 4740