SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue
SpeakerMem-R1 uses dual-track speaker memory to improve multi-party dialogue recall across GroupMem, SocialMem, and EverMem benchmarks.
SpeakerMem-R1 keeps speaker-labeled verbatim messages and derived person- and group-level states, then joins evidence by entity, event, and time. Writer-R1 is trained with SpeakerLevenshtein and speaker-conditioned GRPO to reduce attribution and update errors for local deployment. Binary accuracies are 47.9% on GroupMemBench, 69.2% on SocialMemBench, and 61.9% on EverMemBench, with 62.33% on the public EverMemBench leaderboard. It scores 70.85% on 1,986 LoCoMo questions, and reinforcement learning raises a 305-question set from 57.38% to 68.20%.
- Dual-track memory stores verbatim messages plus person- and group-level states.
- Writer-R1 is trained with SpeakerLevenshtein and speaker-conditioned GRPO.
- Reported accuracies: 47.9% GroupMemBench, 69.2% SocialMemBench, 61.9% EverMemBench.
- Public EverMemBench score is 62.33%; LoCoMo score is 70.85% on 1,986 questions.
- RL lifts a 305-question evaluation from 57.38% to 68.20%.
Full article255 words · extracted from arxiv.org · click to collapse
Long-term conversational memory in multi-party settings requires more than retrieving relevant content from long-term conversations: it must distinguish who said what, whom each statement concerns, how individuals perceive one another, what information is shared by the group, and how states change over time. Recent studies on multi-party dialogue benchmarks show that existing general-purpose LLM memory systems tend to lose person and group relations or struggle to integrate clues distributed across members, groups, and time. Together, these issues reveal two core bottlenecks: message attribution and relational understanding in multi-party dialogue, and state reconstruction from interleaved histories. To address both, we propose $\textbf{SpeakerMem-R1}$: its dual-track memory stores speaker-labeled verbatim messages and derived states organized into person-level and group-level views, then combines evidence from both tracks by entity, event, and time at query time. To reduce attribution and update errors during structured memory construction while enabling local deployment, we train Writer-R1 with SpeakerLevenshtein and speaker-conditioned GRPO. On GroupMemBench, SocialMemBench, and EverMemBench, SpeakerMem-R1 achieves binary accuracies of 47.9%, 69.2%, and 61.9%, respectively. On the publicly reported EverMemBench leaderboard from EverMind-AI, we achieves 62.33%, the best reported result among the latest state-of-the-art frameworks. It also achieves 70.85% on all 1,986 LoCoMo questions, which we use as a two-person long-term conversation boundary test. In a controlled evaluation of 305 questions, RL raises the SFT Writer's mean accuracy from 57.38% to 68.20%. We report both binary accuracy and token-F1, and ablations show that the verbatim and structured tracks, as well as person-level and group-level views, are complementary under the standardized evaluation interface.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.26780