AnalysisNov 17, 20266 min read4 sources

Looki's Loom memory models are real but internal-only, with benchmarks no one else can check

In one minute

The wearable-camera maker describes a stack of three model layers, Sense, Moment and Memory, that turns all-day first-person video and audio into time-stamped memory. Every number it reports comes from Looki's own datasets, and the only public way in is a demo site.

Looki makes the L1, a clip-on camera sold as an AI life-memory device. Its product page promises to "turn everyday moments into searchable memories, daily recaps, and personal insights." In mid-November 2026 Looki AI Research published a Chinese-language article announcing Loom, which it calls a series of in-house multimodal models. Loom has three layers. Sense detects what is happening right now. Moment cuts the continuous stream into events and writes descriptions with timestamps and links to the evidence. Memory rolls those events up into daily, weekly and monthly summaries and long-term memory.

Builders of agents, memory layers and egocentric-video pipelines should care because Looki is attacking a hard, common problem: turning sparse, noisy, always-on first-person data into context an agent can search. The question for builders is whether Loom is something they can use or measure, or just a description of Looki's own product backend.

What we checked

  • Is there a public primary source from Looki documenting Loom, its Sense and Moment models, the reported benchmarks and the access terms? partly confirmed. Looki AI Research published an article announcing Loom with Sense, Moment and Memory layers and benchmark results. The readable copy is a full republication, since the WeChat original sits behind a verification wall. The article points only to a demo at loom.looki.tech, which serves an empty 'Loom' shell. It documents no API, weights, pricing or access terms. [1][4]
  • Can outsiders reproduce Loom's benchmark results? refuted. The Sense scores come from an internal business dataset. Memory was scored on a benchmark Looki built itself. The Moment scores appear only in charts. No dataset, configuration or run link is published. [1]

What we found

  • For Loom Moment, Looki reports Segment F1 above Qwen Omni and Gemini Flash at all three tIoU thresholds (0.5, 0.7 and 0.9), but the scores appear only in charts. For the event descriptions themselves, Looki says Moment beats Qwen Omni on text consistency but still trails Gemini Flash. It claims better evidence attribution. The article does not name model versions. [1]
  • Looki's memory system, which it calls looki memory, scored 79.0 overall. Looki says that is 18 points above mem0 and 15 above A-Mem, while processing only 14.5% and 30.5% as much information per question as those two systems. The comparison ran on a benchmark Looki built from consented, de-identified user recordings, and that benchmark is not published. [1]
  • Looki's English product and blog pages describe the features Loom would power: searchable memories, recaps, local storage with selective sync, and cloud-side context management. None of the pages we opened mentions Loom. The Loom launch exists so far only in Chinese. [2][3]

What it means

Treat Loom as a well-described product backend, not something you can build on. The design is the useful part. Sense uses question-conditioned scoring rather than text generation, so several labels can be true at once, plus a temporal loss that suppresses flicker. Moment detects event boundaries and attaches the source evidence. Memory adds daily, weekly and monthly summaries over the event layer and navigates top-down from summaries to specific moments. Teams building egocentric or always-on agent memory can borrow all three patterns. Every number, though, comes from Looki's own data and unnamed or unversioned baselines. The only number that comes with its operating point is 0.51 s per detection. Even that one has no hardware or batch size attached, and nobody outside Looki has measured any of these figures.

If you are choosing a memory layer, don't read the 79.0 against mem0 and A-Mem as a ranking. Run your own evaluation on your own traces. If you need egocentric segmentation, note that by Looki's own account Gemini Flash still writes more consistent descriptions. There is no API, so the open question for builders is whether Looki ever exposes Loom outside its app.

Competing readings

  • Loom is a real in-house model stack that serves Looki's own wearable product. It is not a platform for outside builders (likely, about 65%). What would tell it apart: Looki announced Loom as its own model series but points only to a demo site, with no API or terms [1][4]; The product page already promises searchable memories and recaps, the job Loom describes [2].
  • Loom is largely a marketing label for a pipeline, and its headline advantages are inflated by self-chosen benchmarks and baselines (unlikely, about 25%). What would tell it apart: Sense was scored on an internal dataset against unnamed references, and the cost claim depends on charted assumptions [1]; The memory benchmark is self-built and unpublished [1].
  • Loom is a precursor to a developer-facing model or API that Looki will open soon (very unlikely, about 10%). What would tell it apart: A dedicated domain, loom.looki.tech, exists for trying the models [1].

Base rate

Of 88 earlier release stories about companies, models, products or projects in the New Runtime archive over the last 12 months that are at least 3 months old, 72 (82%) were followed within 3 months by a access-change or pricing story on the same company or product line. In New Runtime's archive, 72 of 88 earlier release stories were followed within 3 months by an access-change or pricing story on the same product line. That makes some follow-up access news plausible. Loom, however, ships with no API at all, so a follow-up would more likely be a demo or app change than developer pricing.

What would change this view

  • Looki publishes API docs or a pricing page for Loom Sense, Moment or Memory (pricing signal): Loom is a platform for builders, not just a product backend.
  • Looki releases its evaluation sets or model weights publicly (open-source signal): The benchmark claims become checkable, and the reading that they are marketing-inflated weakens.
  • Independent tests show Loom Moment underperforming Gemini Flash or Qwen Omni on segmentation (model-listing signal): The headline segmentation lead does not hold up outside Looki's own data.

Still open

  • Will Looki publish its Sense, Moment or Memory evaluation sets, or the configurations behind them, so the 88.6 F1 and 79.0 scores can be reproduced?
  • Which reference models, versions, hardware and precision sit behind the 0.51 s latency and the roughly 94% cost-reduction claims?
  • Will loom.looki.tech turn into a public demo or API with pricing and terms? We searched for an API or pricing page that would contradict this report and found none.
  • Do L1 users already get Loom in the app, and which features does it replace? Looki's English pages don't say.

Evidence / sources

  1. 1ai.codefather.cn/news/2121269638053556229
  2. 2looki.ai/pages/looki-l1
  3. 3looki.ai/blogs/news/beyond-the-chatbot-how-looki-is-building-proactive-ai
  4. 4loom.looki.tech
All Analysis

Next reads

Read next

Open the graph
  1. 01topicMemoryExplore the memory topic hub.
  2. 02topicGeminiExplore the gemini topic hub.
  3. 03topicMultimodalExplore the multimodal topic hub.
  4. 04archiveAnalysisOpen the latest authorial analysis.
  5. 05sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract