VoxPolyMem: Interaction-aware multimodal memory for multi-party spoken conversations beats baseline by 23.6 points

Wenxu Jia · hf · 2026-09-30

VoxPolyMem tackles long-term memory for multi-party spoken conversations, combining incremental speaker identification with a memory hierarchy of interaction memory, fact memory, and participant profiles. Retrieval is framed as sequential decision-making with an agent rewriting queries and selecting tools, trained via Evidence-Gain GRPO (EG-GRPO) for round-wise credit assignment. The authors also release VoxPolyBench. It scores 85.0 overall, beating the strongest baseline by 23.6 points, and exceeds public memory baselines by 8+ points on Mem-Gallery and H2HMem-Multi. Code and datasets are open-sourced.

Original post →

More from coding & agent

coding & agent channel →