HGP:An on-device personalized agent memory via hybrid graph storage

2026-10-07

AI Summary

This study addresses the challenges of inaccurate retrieval and deployment difficulties in on-device personalized interactions for LLM-based agents, which arise from heterogeneous implicit constraints within long-term memory. To this end, we propose a hybrid graph storage framework that structures episodic, semantic, and procedural memories into graphs, leveraging working memory trajectories to capture states and enable precise routing. Furthermore, we introduce a novel lightweight self-enhancing classifier that overcomes the limitations of single-vector representations and substantially reduces large model invocations, thereby facilitating efficient edge deployment. Evaluated on the PAL-Set benchmark, the proposed approach achieves a solution selection score of 35.58, outperforming the strongest baseline by nearly seven points and significantly enhancing the reliability of personalized decision-making.

Abstract

LLM-based agents face challenges in personalized interactive tasks due to heterogeneous, multi-typed, and implicitly constrained long-term traces. Existing memory mechanisms struggle with accurate routing and retrieval, especially on-device where personalization is critical. Most methods use single-vector representations, blurring type distinctions and relational structure. We propose HGP, a hybrid graph memory framework. HGP employs a lightweight self-enhancement classifier for personalized memory routing and constructs episodic, semantic, and procedural memories as graphs. It also extracts working memory as a state trajectory to capture current state and implicit constraints, ensuring reliable decision-making. The classifier reduces large-model calls, enabling on-device deployment, while graph storage enables accurate retrieval and incremental user profile refinement. Experiments on two benchmarks show that on PAL-Set solution selection, HGP achieves an S-score of 35.58, nearly 7 points above the strongest baseline. Code and data are at https://github.com/Ouan6/HGP-.git.

Problem

• LLM-based agents
• personalized memory
• on-device deployment
• memory retrieval
• hybrid graph storage

Innovation

• Hybrid Graph Memory
• On-device Deployment
• Personalized Routing
• State Trajectory
• Lightweight Classifier