SpeakerMem-R1 posts gains on four multi-party memory benchmarks

Multi-party conversation memory has a specific failure mode. In a group chat with five participants, who said what and whom each remark concerns are both live variables, and general-purpose LLM memory systems, per a new arxiv paper from Haobo Zheng and colleagues, "tend to lose person and group relations."

SpeakerMem-R1's fix is a dual-track store. One track keeps speaker-labeled verbatim messages; the other keeps derived states, organized into person-level and group-level views. At query time the system combines evidence from both by entity, event, and time. Memory, in this framing, "must distinguish who said what, whom each statement concerns, and how individuals perceive one another."

The reported numbers span four benchmarks. GroupMemBench: 47.9% binary accuracy. SocialMemBench: 69.2%. EverMemBench: 61.9%, with a 62.33% leaderboard entry listed as EverMind-AI. LoCoMo: 70.85% across all 1,986 questions.

A separate controlled evaluation on 305 questions isolates the reinforcement-learning step, a Writer-R1 trained with "SpeakerLevenshtein and speaker-conditioned GRPO," and reports mean accuracy climbing from 57.38% to 68.20%, roughly an eleven-point gain on that slice.

The retrieved abstract publishes those four top-line accuracies and the RL delta, but not baseline scores from prior memory systems on the same benchmarks, so the paper's headline claim stands without a side-by-side comparator in the abstract itself.