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

What it claims

The page attaches concrete numbers the summary omits. On the three target benchmarks, SpeakerMem-R1 reports binary accuracies of 47.9% (GroupMemBench), 69.2% (SocialMemBench), and 61.9% (EverMemBench). On the publicly reported EverMemBench leaderboard from EverMind-AI it reaches 62.33%, which the authors call the best among current SOTA frameworks. On all 1,986 LoCoMo questions it scores 70.85%, using LoCoMo as a two-person "boundary" case for its multi-party design.

How it works

Two named bottlenecks motivate the design: (1) message attribution plus relational understanding, and (2) state reconstruction from interleaved histories. The dual track stores speaker-labeled verbatim messages and derived states organized into person-level and group-level views, then merges evidence from both tracks by entity, event, and time at query time. The Writer-R1 model that builds the structured memory is trained with SpeakerLevenshtein (a speaker-aware edit metric) and speaker-conditioned GRPO, explicitly to enable local deployment rather than relying on a hosted model.

Evidence

The RL result is the sharpest data point the page gives: in a controlled evaluation of 305 questions, RL raises the SFT Writer's mean accuracy from 57.38% to 68.20% — a ~11-point gain attributable to the GRPO stage over supervised fine-tuning. The authors report both binary accuracy and token-F1. Ablations are stated to show that the verbatim and structured tracks, and the person- and group-level views, are complementary under a standardized evaluation interface (i.e. each contributes; neither alone matches the full system). Exact ablation deltas are not in the abstract.

Limits

This is an abstract-only page — no code, weights, licence, model size, or training-compute figures appear, so "enables local deployment" is a claim without a stated model scale to check. The strongest benchmark numbers are still low in absolute terms (47.9% binary accuracy on GroupMemBench), so these are hard tasks, not solved ones. The LoCoMo result is explicitly a two-person degenerate case of the multi-party method, not the target regime. A v2 exists (23 Sep 2026); the numbers above are from v1.