openJiuwen AgentMemory 0.2.0 adds OmniMem-Lens for event-based long-video memory
Overview
openJiuwen AgentMemory 0.2.0 launches OmniMem-Lens, a multimodal memory capability for long videos, according to QbitAI (量子位).
Instead of cutting video into fixed one-minute windows, it segments footage by real events, using visual and speech (ASR) streams. The memory is organized in three layers and stored as text and vectors that can connect to mainstream vector databases.
In tests against Qwen3-VL-8B, the project reports recall accuracy up to 45.7% higher on EgoLife-QA and online question-answer latency up to 97% lower. On LVbench, response time was 1.72 seconds versus 34.5 seconds for direct VLM answering. These are the project's own test results as reported.
WorkSwarm says it will soon support connecting to openJiuwen AgentMemory to store video content from daily conversations and help users recall it during discussions.
Written by AI from the articles below · updated Nov 3, 11:04 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- QbitAINewsopenJiuwen AgentMemory adds OmniMem-Lens for event-based long video memory
AIopenJiuwen AgentMemory 0.2.0 launches OmniMem-Lens, a multimodal memory capability that segments long videos by real events using visual and speech (ASR) streams instead of fixed one-minute windows. In tests against Qwen3-VL-8B, it raises memory recall accuracy by up to 45.7% on EgoLife-QA and cuts online question-answer latency by up to 97%, with a 1.72-second response on LVbench versus 34.5 seconds for direct VLM answering. The memory is organized in three layers and stored as text and vectors that can connect to mainstream vector databases.
Heat Trend
Not enough continuous observations to show a trend yet.