Jiaxin Ran’s paper on graph-based dialog memory, titled “Does Memory Need Graphs? A Unified Framework and Empirical Analysis for Long-Term Dialog Memory”, has been accepted to ACL 2026.

Graph structures are increasingly adopted in dialog memory systems, motivated by their success in retrieval-augmented generation and the associative nature of human memory. However, empirical findings on their effectiveness remain inconsistent, making it unclear which design choices truly matter. To address this, we propose an experimental and system-oriented analysis of long-term dialog memory architectures. We formalize 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 the LongMemEval and HaluMem benchmarks, comparing common design choices in memory representation, organization and maintenance, as well as indexing and retrieval. Our results show that underlying implementation details—often insufficiently specified in prior work—have a substantial impact on performance. Furthermore, we identify stable, reliable strong baselines to support fair comparison and practical deployment.