Title: RGMem: Renormalization Group-Inspired Memory Evolution for Language Agents

Abstract:

For LLM-based conversational agents, maintaining personalized and continuous interactions is essential. However, the limitations of finite context windows and static parametric memory often obstruct the accurate modeling of long-term, cross-session user states. Current methods, such as retrieval-augmented generation and explicit memory systems, largely function at the level of individual facts. This approach struggles to extract stable preferences and deep-seated user traits from dialogues that are constantly evolving and may contain contradictions.

To overcome these limitations, we introduce RGMem, a self-evolving memory framework grounded in the renormalization group (RG) perspective on multi-scale organization and emergence. RGMem conceptualizes long-term conversational memory as a multi-scale evolutionary process. In this framework, episodic interactions are first converted into semantic facts and user insights. These elements are then progressively integrated via hierarchical coarse-graining, thresholded updates, and rescaling, resulting in a dynamically evolving user profile.

By distinctly separating rapidly changing evidence from slowly varying traits, and by facilitating non-linear, phase-transition-like dynamics, RGMem facilitates robust personalization that surpasses the capabilities of flat retrieval or static summarization techniques. Comprehensive experiments conducted on the LOCOMO and PersonaMem benchmarks show that RGMem consistently outperforms state-of-the-art memory systems, delivering superior cross-session continuity and enhanced adaptation to shifting user preferences. The code for this project is accessible at https://github.com/fenhg297/RGMem.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC