Looki's Loom gives a wearable camera a memory that scores 79, 18 points above mem0
A wearable camera records hours of video and audio a day, and nobody will ever rewatch it. Looki, the maker of the L1 camera, built Loom: the system splits recordings into events on its own and answers questions like "what did we agree on with the supplier?" while showing the source frames. On Looki's own test, Loom scored 79, which is 18 points above mem0 and 15 above A-Mem. No independent lab has verified the results yet.
AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
A wearable camera records hours of video and audio a day, and nobody is going to rewatch them. To make the footage useful, an assistant has to answer a question like "what did we agree on with the supplier last week?" and show the frames the answer came from.
That is what Loom does. It comes from Looki, the Chinese maker of the Looki L1 wearable camera. On the company's own test, built from real user recordings collected with their consent, Loom's memory scored 79.
That is 18 points above mem0 and 15 above A-Mem. Both competitors are popular memory systems for text-based AI agents, meaning they handle conversations and documents, not video. The practical takeaway for developers: if you are building an assistant with a camera or microphone, the "events — summaries — source frames" scheme looks like a workable alternative to flat text memory, but so far only Looki itself has produced the numbers.
What Loom is made of
The system consists of two models and a memory layer.
Sense answers the question "what is happening right now." It receives a short clip of video and audio, a question about the event, and several answer options. The model does not write a text description. Instead, it scores each option. For an always-on device this makes sense: picking from a list is faster and cheaper than generating text. According to Looki, on an internal set of nearly 20 event types, Sense reaches an F1 of 88.6 (a general measure of how accurately and completely a model finds the target events) with an average recognition time of 0.51 seconds.
Moment works with material that has already been recorded. It aligns picture and sound on a single timeline, finds event boundaries on its own, and describes the events. Each event has a start and end time, a description, and supporting evidence. Next to every fact is a label saying whether it was seen, heard, or inferred from context. This labeling matters for trust: users can tell what actually appeared in the frame from what is the model's guess. Looki claims that Moment identifies event boundaries more accurately than Qwen Omni and Gemini Flash. The source gives no specific figures.
How the memory works
The Memory layer assembles events into summaries for a day, a week, and a month. It also builds an overall snapshot of what the system knows about the user: habits, surroundings, recurring tasks.
When the user asks a question, the system works top-down. First it looks at the monthly summaries, then drills down to specific events and, if needed, to the source frames. It is like searching a paper archive: you open the right folder first, then the right sheet, and then check the original document. The model does not have to hold a whole month of recordings in context, and the answer can be checked against the primary source.
The gap with mem0 and A-Mem looks predictable: both systems were built for text, and video and audio from a camera are a different kind of material.
What to do with this, and the caveats
For anyone building wearables, voice and video assistants, or a "second memory" for users, Loom offers a ready-made blueprint:
- a separate fast model for event recognition;
- a separate step that cuts the stream into episodes and cites the source of each fact;
- a hierarchy of summaries, where search runs from the general to the specific and ends with a link to the original frame.
The caveats are significant. All figures come from Looki's own internal datasets, and there is no independent verification. The comparison with mem0 and A-Mem was run on a benchmark the company assembled itself, so outside its data the gap may turn out different. For Moment there are no numbers, only the word "more accurate." Finally, constant recording of the people around the wearer raises privacy questions. The source says only that the test data was collected with users' consent. It does not specify how the system handles other people's voices and faces.
Access to Loom is limited. You can try it in a web demo and there is an API, but Looki does not plan to release the system as open source. Third-party developers can either use the API or reproduce the architecture themselves.
The company describes its next step this way: merge all three parts into a single multimodal model that decides on its own when to offer help. The assistant would then stop merely answering questions and start stepping in on its own initiative. That is harder, because the model has to understand not only what is happening but also whether it is appropriate to speak right now. For now, Loom shows that memory for a wearable camera is worth designing as a separate layer with links back to the originals.
- Nov 18, 2026In 94% of B2B shortlists, AI summaries barely show the companies' own websites
- Nov 18, 2026A Video From One Product Photo Cost $4.25 Instead of $2.42 Because of Seedance 2.0 Defects
- Nov 18, 2026Claude Code saved passwords from commands: secrets found in 1.1% of settings files on GitHub
- Nov 18, 2026OpenAI releases GPT-6.2 Astra: second place in the rankings and a 20% lower price
Key Questions Answered
What Loom is made of?
The system consists of two models and a memory layer. Sense answers the question "what is happening right now." It receives a short clip of video and audio, a question about the event, and several answer options. The model does not write a text description. Instead, it scores each option. For an always-on device this makes sense: picking from a list is faster and cheaper than generating text.
How the memory works?
The Memory layer assembles events into summaries for a day, a week, and a month. It also builds an overall snapshot of what the system knows about the user: habits, surroundings, recurring tasks. When the user asks a question, the system works top-down. First it looks at the monthly summaries, then drills down to specific events and, if needed, to the source frames. It is like searching a paper archive: you open the right folder first, then the right sheet, and then check the original document.
What to do with this, and the caveats?
For anyone building wearables, voice and video assistants, or a "second memory" for users, Loom offers a ready-made blueprint: a separate fast model for event recognition; a separate step that cuts the stream into episodes and cites the source of each fact; a hierarchy of summaries, where search runs from the general to the specific and ends with a link to the original frame. The caveats are significant. All figures come from Looki's own internal datasets, and there is no independent verification.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.