VoiceMem brings streaming memory to voice agents

VoiceMem makes a strong case that voice-agent memory needs both context discipline and emotional continuity, though its benchmark claims still need independent reproduction.

• –Dual-brain storage separates durable facts from evolving feelings, preferences, and relationships
• –Streaming retrieval can hide memory latency inside the user’s speaking turn
• –Top-five retrieval and roughly 430 memory tokens point toward leaner, more controllable agent contexts
• –Multimodal ingestion covers speech, speakers, sound events, and multi-party conversations
• –The modular, open-source design should make it easier to swap memory engines and voice models