SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue

Dual-track speaker memory with an RL-trained Writer-R1 reaches 62.33% on the public EverMemBench leaderboard and 70.85% on 1,986 LoCoMo questions.

SpeakerMem-R1 is a dual-track long-term memory system for multi-party dialogue that stores speaker-labeled verbatim messages alongside derived states organized into person-level and group-level views, joined by entity, event, and time. Its Writer-R1 model is trained with a SpeakerLevenshtein loss and speaker-conditioned GRPO to reduce attribution and update errors; the arXiv report notes this enables local deployment. Reported binary accuracies are 47.9% on GroupMemBench, 69.2% on SocialMemBench, and 61.9% on EverMemBench, alongside 62.33% on the public EverMemBench leaderboard (identified as EverMind-AI's leaderboard in the Hugging Face report, which says it leads reported results) and 70.85% on 1,986 LoCoMo questions. On 305 controlled questions, reinforcement learning raises the supervised Writer's mean accuracy from 57.38% to 68.20%. The paper surfaced on Hugging Face daily papers on 2026-09-21 and was listed on arXiv (cs.AI, cs.CL, cs.LG) on 2026-09-22; the two reports agree on all figures.

• Dual-track memory stores speaker-labeled verbatim messages plus derived person-level and group-level states, combined by entity, event, and time.
• Writer-R1 is trained with SpeakerLevenshtein and speaker-conditioned GRPO to reduce attribution and update errors; the arXiv report says this supports local deployment.
• Binary accuracies: 47.9% GroupMemBench, 69.2% SocialMemBench, 61.9% EverMemBench.
• 62.33% on the public EverMemBench leaderboard (EverMind-AI's leaderboard per the Hugging Face report), which that report says leads reported results.
• 70.85% on 1,986 LoCoMo questions.
• Reinforcement learning raises the supervised Writer's mean accuracy on 305 controlled questions from 57.38% to 68.20%.
• Listed on Hugging Face daily papers 2026-09-21 and on arXiv (cs.AI, cs.CL, cs.LG) 2026-09-22; the two reports do not conflict.