Looki launches life understanding model Loom: perception, experience, and long-term memory processed in three layers.

Beating AI News Flash: AI wearable device company Looki has released Loom, a multimodal model series that can organize daily life recorded by cameras and microphones into long-term memories with timestamps, content, and raw evidence, retrievable by AI at any time.

For example, if a user has lunch with colleagues and discusses a project during the meal, ultimately deciding to wait until Friday's data comes out before making a decision, Loom must not only recognize "eating" and "conversing" but also record the discussion content, the decision, and the reasons behind it. When the project comes up again days later, the AI can retrieve the records from that time.

Loom is divided into three parts:

Sense (Perception): Continuously determines what the user is doing. It uses specially trained discriminative models that directly score questions like "is the user eating" and "is the user conversing" without generating text descriptions each time. The model also references the previous time period's state to avoid activity recognition from flip-flopping due to turning around or occlusion.

Moment (Experience): Aligns visuals and audio collected at different times, automatically determines where an experience begins and ends, and generates descriptions containing key facts, timestamps, and corresponding evidence. It can segment experiences based on conversation content rather than simply cutting video by fixed duration.

Memory: Organizes experiences into daily, weekly, and monthly summaries as well as long-term memories. When answering questions, the AI can first check summaries, then drill down layer by layer to specific experiences, and when necessary review the original footage or recordings to verify details.

Currently, Loom still operates through multiple specialized models working in coordination with a memory retrieval system, with plans to further integrate these capabilities into an end-to-end multimodal model in the future.

In official self-testing, Sense recognized nearly 20 categories of daily activities with an F1 score of 88.6 and an average judgment time of 0.51 seconds per instance. Moment surpassed Gemini Flash in determining the start and end times of experiences, but its generated text descriptions are still less accurate than the latter. Memory scored 79.0, 18 points higher than mem0, with average Token processing per question only 14.5% of the latter's.

Loom will first be used in Looki's own wearable devices. Currently, no standalone model API is available, and the model weights have not been open-sourced.