VoiceMem is a memory architecture for duplex speech language models built around a parallel ‘informational left brain’ and ’emotional right brain,’ paired with streaming memory read/write mechanisms designed for real-time conversation. The system outperforms memory baselines such as Mem0 by nearly 30 points under top-5 retrieval and improves on state-of-the-art persona benchmarks by 4.29 points via dual-node persona modeling. In deployment, VoiceMem completes memory retrieval in 134ms, within standard voice-activity-detection latency budgets and without added conversational delay.
