openJiuwen AgentMemory 0.2.0 Adds OmniMem-Lens for Event-Level Long-Video Memory

openJiuwen has released OmniMem-Lens as part of AgentMemory 0.2.0, a multimodal memory capability aimed at a concrete problem: how an agent should remember a three-hour video so it can retrieve the right moment when asked. According to figures published with the release, the approach lifts memory recall by up to 45.7% and cuts online question-answering latency by up to 97%.

The bottleneck is familiar. Long videos quickly exceed an agent's context window, and both throughput and accuracy degrade as footage piles up. The common workaround is to slice the video into fixed windows, often one minute each, and let the agent memorize the pieces separately. That keeps the pipeline simple but breaks continuous content into unrelated fragments.

The team describes two failure modes. First, evidence gets fragmented: a single explanation is cut at a window boundary, so the agent retrieves half an argument and hallucinates the rest. Second, evidence gets diluted: a few seconds of critical action are bundled with minutes of irrelevant footage, and the detail is buried when the agent searches for it.

OmniMem-Lens tracks two streams at once. The visual stream detects scene changes and marks them as boundaries; the ASR stream catches moments when the picture stays still but the speaker moves to a new topic. Cross-checking the two signals places cuts at real event boundaries rather than at stopwatch intervals.

Memory is then organized in three layers. L1 works like a table of contents, L2 like an index that points to specific content, and L0 keeps the original frames for verification. L1 and L2 units are fully textualized and vectorized, so they can plug into mainstream vector databases with little change to an existing RAG or agent stack.

The heavier work happens offline. Memory is built ahead of time, and at query time the system matches the question against L1 and L2 units first, pulling precise L0 frames only when visual verification is actually needed. That lightweight-text-index-plus-deferred-video-retrieval path avoids re-watching hours of footage for every question.

The team benchmarked the approach against a Qwen3-VL-8B model answering video questions directly. Recall rose 8% on Video-MME-L, 6.4% on LVbench, and 45.7% on EgoLife-QA, an egocentric dataset that records a full day of daily life. On LVbench, direct answering took 34.5 seconds while the OmniMem-Lens path took 1.72 seconds, roughly 20 times faster.

Two examples illustrate the gap. In a two-hour animated film, a sequence running from 14:20 to 16:50, in which a character leaves, is watched, and is then imagined as a butterfly, spans three one-minute windows and loses its causal thread when cut twice. A 40-minute product review with a static architecture diagram is the opposite case: the speaker covers architecture, load testing, and next quarter's schedule, so vision alone merges three topics into one memory while ASR separates them.

When the unit of memory shifts from a fixed time window to a real event, both accuracy and response time improve at once. The team frames OmniMem-Lens as a super hippocampus for agents; it is live in AgentMemory 0.2.0, with code hosted in the openJiuwen/agent-memory repository on Gitcode, and WorkSwarm is slated to integrate the capability. For the agent stack, long-video memory is becoming less a question of whether footage can be stored and more a question of whether it is stored and retrieved correctly.