openJiuwen releases OmniMem-Lens to split long videos at real events, cutting latency up to 97%

openJiuwen, the Huawei-backed open-source agent framework, has released OmniMem-Lens in version 0.2.0 of its AgentMemory module. The multimodal memory capability lets AI agents remember long videos by real events rather than fixed one-minute windows, lifting recall up to 45.7% and cutting answer times by up to 97%, QbitAI reported on Pandaily.

The team argues that the common approach—cutting long videos into one-minute windows and storing each separately—breaks agent memory in two ways. A continuous scene split across windows leaves the agent with half the evidence, inviting hallucination. A few seconds of key action bundled with minutes of irrelevant footage gets buried in noise, lowering retrieval accuracy. Most event-segmentation methods also rely on a single modality, missing topic changes when the picture stays still or scene changes when nobody speaks.

OmniMem-Lens tracks both streams. A visual track flags scene changes, while an audio track uses speech recognition to detect when a speaker switches topic even though the screen has not changed. Cross-checking the two streams cuts at real event boundaries. In one example, a two-hour product review shows the same architecture diagram for 40 minutes while the speaker covers three separate topics; the audio track splits them into three memories so a later question about load-test figures can be answered.

Memory is stored in three layers: L0 keeps the raw video, L1 holds event-level memories and L2 holds finer segment-level memories. L1 and L2 are fully converted to text and vectors, so they can plug into mainstream vector databases and existing retrieval-augmented generation stacks with little rework. The heavy organizing work is done offline. At query time the system matches the query against L1 and L2, and only pulls exact frames from L0 when visual details need checking.

The team tested OmniMem-Lens paired with Qwen3-VL-8B against Qwen3-VL-8B answering directly from the video. Recall accuracy rose 8% on Video-MME-L and 6.4% on LVBench. The largest gain, 45.7%, came on EgoLife-QA, a first-person dataset recording a full day of daily life, where fixed slicing tends to fail. On LVBench, online response time fell from 34.5 seconds to 1.72 seconds, about 20 times faster, and the team puts the maximum latency reduction at 97%.

The project code is available on GitCode with a mirror on GitHub under the Apache-2.0 license. The team says openJiuwen WorkSwarm will soon support AgentMemory, so video shared in everyday conversations can be remembered and recalled during later discussions.