Year: 2026
Type: conference-paper
Abstract: Integrating knowledge graphs (KGs) with large language models (LLMs) enhances factual accuracy and interpretability in question answering (QA).However, existing agent-based methods rely on static memory mechanisms, which fail to address the state-space explosion in multi-hop reasoning and lack self-evolution capabilities.To overcome these limitations, we propose EvoMemKG, an agent framework with an evolvable memory mechanism for multi-hop KG reasoning.EvoMemKG introduces a dual-layer memory: (1) a working memory that manages and compresses intermediate states using a relation-based compression algorithm, effectively reducing context usage; and (2) an experience memory that distills historical reasoning paths into reusable, generalized strategies, enabling cross-task knowledge transfer and self-evolution.We further design a double-loop workflow that orchestrates the LLM, the dual-layer memory, and KG environment to enable end-to-end autonomous reasoning.Extensive evaluations on three KGQA datasets across two KGs demonstrate that EvoMemKG achieves state-of-the-art performance without requiring additional training or specialized tools, with improvements of up to 20% over strong baselines on complex multihop queries.
Institution: Beijing University of Posts and Telecommunications
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166818039