Chinese company Looki, which produces the Looki L1 wearable camera, has introduced the Loom family of multimodal models. The system converts video and audio recorded throughout the day into structured long-term memory tied to timestamps and original frames. The company reports significant token savings and a latency of about half a second for the perception module, but has so far only opened a web demo — the system has no API or open weights.

image
image
image

What Happened

Information about the Loom launch was published in the Looki AI Research WeChat channel. An analysis dated November 14, 2026, appeared in AI Daily Insights digest No. 102, and a brief description was also published by Anue/BlockBeats. The system is described as having three layers. Loom Sense is a discriminative model that answers natural language questions about the current event without generating a text description: features like "food" and "conversation" can be simultaneously true. Loom Moment simultaneously marks event boundaries and describes them on a common timeline, distinguishing between what is seen, what is heard, and what is inferred from context. Loom Memory stores a hierarchy of "events, then daily, weekly, and monthly summaries, then a user profile snapshot" and, when responding, descends from summaries to specific events and original frames. According to Looki, Sense achieves an F1 of 88.6 with a latency of 0.51 s and approximately 94% savings in inference costs compared to "closed reference models." Moment outperforms Qwen Omni and Gemini Flash in Segment F1@tIoU, while Memory scores 79.0 compared to 61.0 for mem0 and 64.0 for A-Mem, consuming 14.5% of mem0's tokens and 30.5% of A-Mem's tokens. The system is currently available only as a web demo at loom.looki.tech.

Context

This development addresses a typical weakness of wearable AI devices and agent frameworks: memory is often stored as abstract vectors, making it impossible to show where the assistant derived a specific answer — this is the case for mem0 and A-Mem, among others. Loom contrasts this with evidence-based memory: every inference must rely on a frame or audio clip and a note indicating whether the information was seen, heard, or inferred. The solution is structured as a vertical product: the Looki L1 camera, weighing 32 g with a 12 MP sensor, three microphones, 32 GB of storage, and IP67 protection, costs $199 and is always with the owner, while Loom converts its video and audio stream into structured events. The discriminative Sense module is a conscious trade-off between quality and cost: the system classifies an event based on a question rather than describing every fragment in text, which explains the claimed latency and savings. The Memory hierarchy, descending from summaries to frames, addresses a second fundamental pain point for personal assistants: answering questions does not require running raw recordings through the model in their entirety.

Why This Matters for the Industry

For AI product builders, the main value is not the weights but a reproducible pattern: a discriminative event trigger, event labeling on a common timeline with "seen, heard, inferred" tags, and hierarchical memory with references to original frames. This stack promises cheap, traceable memory: the company estimates Loom Memory's token consumption to be approximately 6.9 times lower than mem0's, directly reducing the cost of an "always-on" personal assistant. If the claimed figures hold up to independent verification, hierarchical memory with tracing back to original frames could become the de facto standard for personal assistants and the proactive wearables category — for much the same reasons RAG became the standard for corporate search. Competition would then shift to the quality of summaries and the "show original" interface. In the six-month horizon, much depends on whether Looki publishes a technical report or at least a limited API: this would enable third-party measurements of Segment F1@tIoU and latency, and likely open replications of the pattern; if no publications follow, Loom will remain a marketing showcase for the camera.

Why This Matters for Users

If you use a wearable camera like the Looki L1 or are just choosing one, you can try the system right now: the web demo at loom.looki.tech shows the quality of event labeling and memory responses, and until an API is available, manually checking behavior is the only way to evaluate it. For those working with agent memories like mem0 or A-Mem, Loom provides a clear architectural reference where any answer can be expanded to the original frame or audio, rather than an abstract vector. The figures from Looki's announcement should be included in a checklist when comparing similar systems — but as manufacturer hypotheses, not as independent benchmark measurements. A separate nuance concerns privacy: the proprietary benchmark is built on the company's internal data and Looki user recordings, which, according to the manufacturer, were processed with user consent.

What Is Still Unknown / Limitations

The claimed results cannot be verified: available materials are insufficient to reproduce the models or repeat the figures — there is no technical report, event boundary labeling methodology, or baseline configurations. A key part of the inference comparison relies on unnamed "closed reference models," so the claimed approximately 94% cost savings cannot be independently verified. The memory benchmark is proprietary and built on data from Looki, so the comparison of 79.0 versus 61.0 for mem0 and 64.0 for A-Mem remains a manufacturer claim, and Moment's superiority over Qwen Omni and Gemini Flash in Segment F1@tIoU is neither refuted nor confirmed. There is no separate API, open weights are not planned, and service pricing has not been announced. If a technical report or API does not appear, the benchmarks will remain promotional, and comparable hierarchical memory will have to be built in one's own stack.

Sources

Author

Look at AI, editorial team