Does Memory Need Graphs? A Unified Framework and Empirical Analysis for Long-Term Dialog Memory

TL;DR

This work tackles conflicting findings about graph-based long-term dialog memory by introducing a unified, system-oriented framework that accommodates both graph-based and flat memory designs. Through controlled experiments on LongMemEval and HaluMem, it analyzes memory extraction, indexing, maintenance, and retrieval, uncovering that foundational system settings often drive performance more than architectural novelty. The study identifies stable baselines and graph constructions (e.g., DescGraph) that improve retrieval and QA under certain conditions, while noting that graph memory can incur higher latency and maintenance costs. The results offer practical guidance for building robust, scalable dialog memory systems and establish benchmarks for fair comparison across methods. Overall, the paper emphasizes careful alignment of design choices and presents a set of baselines to advance future dialog-memory research.

Abstract

Graph structures are increasingly used in dialog memory systems, but empirical findings on their effectiveness remain inconsistent, making it unclear which design choices truly matter. We present an experimental, system-oriented analysis of long-term dialog memory architectures. We introduce a unified framework that decomposes dialog memory systems into core components and supports both graph-based and non-graph approaches. Under this framework, we conduct controlled, stage-wise experiments on LongMemEval and HaluMem, comparing common design choices in memory representation, organization, maintenance, and retrieval. Our results show that many performance differences are driven by foundational system settings rather than specific architectural innovations. Based on these findings, we identify stable and reliable strong baselines for future dialog memory research.

Figures (1)

• Figure 1: The unified framework for dialogue memory system.