Giving Agents "Audio-Visual Dual-Stream Memory" for the First Time: openJiuwen Agent-Memory Understands Long Videos Without Missing a Single Key Event

What happened: openJiuwen AgentMemory 0.2.0 launched OmniMem-Lens, a multimodal memory capability that segments long videos by real events using combined visual and audio streams instead of fixed time windows. Key numbers: Recall precision up to 45.7% higher (EgoLife-QA), +8% on Video-MME-L, +6.4% on LVbench. Why it matters: The team says fixed-window splitting fragments and dilutes evidence, causing hallucination and missed details. OmniMem-Lens builds memory offline in three layers and retrieves online, which the content presents as enabling faster, more accurate long-video QA.

openJiuwen AgentMemory 0.2.0 launched OmniMem-Lens, a multimodal memory capability that segments long videos by real events using combined visual and audio streams instead of fixed time windows.

The team says fixed-window splitting fragments and dilutes evidence, causing hallucination and missed details. OmniMem-Lens builds memory offline in three layers and retrieves online, which the content presents as enabling faster, more accurate long-video QA.