2026-11-13 15:39

Introduction: Weaving first-person life into verifiable, long-term memory. What is Loom? Loom is a multimodal model series launched by Looki, designed to understand first-person lived experiences.

What is Loom?

Loom is a multimodal model series introduced by Looki for understanding first-person life. Starting from visual and audio data collected by devices, Loom Sense identifies what is happening in real time; Loom Moment stitches fragmented observations into traceable experiences with start/end timestamps and evidence; Loom Memory enables hierarchical storage and cross-temporal retrieval of memories. Together, these components empower AI to proactively assist at the right moment, allowing users to focus on living rather than managing their lives.

Core Functions of Loom

• Loom Sense (Event Perception): Based on continuous visual, audio, temporal inputs, and prior state, it determines “what is currently happening,” delivering stable, dynamically updated event states.
• Loom Moment (Experience Structuring): Organizes continuous multimodal observations into structured experience segments with start/end timestamps, content descriptions, and verifiable evidence.
• Loom Memory (Memory Retrieval): Leverages hierarchical storage and tree-based navigation to retrieve past experiences and apply them contextually when answering new queries.
• Proactive Intelligence: Integrates current context and user intent to determine when and how to intervene.

Technical Principles of Loom

• Loom Sense: Conditioned event classification inspired by JEV-style frameworks. Events are defined via natural language questions. The model directly outputs candidate scores based on visual, audio, temporal cues, and previous state. Scores are normalized via Softmax within the candidate set, ensuring independent event judgments while reusing shared encoders to control computational overhead. Temporal consistency is enhanced through state propagation and dual-constraint loss: suppressing probability jitter during ongoing activities and strengthening switching supervision near true boundaries—balancing stability and responsiveness.
• Loom Moment: Joint generation of boundaries and descriptions. Unifies “where to split” and “what each segment contains” into a single generation task. The model simultaneously processes multimodal observations, real-time metadata, and contextual information, outputting start/end times, content descriptions, and supporting evidence in one pass—eliminating error accumulation from sequential segmentation and description. Training proceeds in two stages: first learning boundaries and event types, then adding supervision for content and evidence attribution. Time is embedded as structural input, distinguishing between observed, heard, and inferred elements to prevent hallucination.
• Loom Memory: Hierarchical storage and agentic memory navigation. Long-term memory is organized into four layers: Experience → Summary → Pending Integration → Snapshot. Snapshots are continuously updated via conflict resolution and forgetting mechanisms, while original experiences remain preserved for verification. Retrieval employs Agentic Memory Navigation, following a tree-like path from high-level overview to granular detail, dynamically adjusting branches based on intermediate results to achieve target answers with minimal record access.
• Proactive Intelligence: End-to-end joint modeling. Text, images, video, audio, and retrieved historical evidence are fed into the model simultaneously, producing contextual understanding, action selection, assistance content, and evidence citations. Results from Sense/Moment/Memory serve as auxiliary signals to constrain the final task. Clear division of labor is maintained: the model layer handles perception and decision-making, while the system layer manages storage and permission isolation.

How to Use Loom

• Access Point: Visit https://loom.looki.tech/ to register and experience Loom’s model capabilities.
• Wear & Authorize Devices: Capture visual and audio data via first-person wearable devices and enable continuous perception within authorized scope.
• Automatic Event Sensing: Loom Sense automatically detects “what is happening now” without manual logging or tagging.
• Auto-generated Experiences: Loom Moment automatically organizes continuous observations into life segments with timestamps, descriptions, and evidence.
• Build Long-term Memory: The system automatically accumulates memories across layers—experiences, summaries, snapshots—and continuously updates them based on user feedback.
• Query Anytime for Recall: Ask natural language questions in the app; Loom Memory retrieves relevant fragments along the timeline to answer.
• Await Proactive Assistance: AI proactively reminds or suggests based on current context and historical memory at optimal moments.

Core Advantages of Loom

• End-to-End Technical Pipeline: From event perception (Sense), to experience structuring (Moment), to memory retrieval (Memory), it covers the full closed-loop lifecycle of “recording → understanding → recalling.”
• Continuous and Stable Event Detection: State propagation and dual-constraint loss suppress false positives caused by occlusion or visual flicker, ensuring smooth transitions while maintaining timely response to real-world changes.
• Evidence-Based, Traceable Memory: Experience segments include timestamps, descriptions, and evidence anchors, with strict distinction between “observed,” “heard,” and “inferred”—enabling verifiable answers.
• Efficient and Low-Cost Retrieval: Shared encoding ensures linearly scalable computation across multiple events. In long-sequence memory benchmarks, Loom processes only 14.5% and 30.5% of the information volume compared to mem0 and A-Mem, respectively.
• Superior Performance: Achieves F1 score of 88.6 in event detection, outperforming both open-source and closed-source baselines. Inference cost is reduced by ~94% versus closed-source alternatives. Memory integration score reaches 79.0—18 points higher than mem0 and 15 points higher than A-Mem.

Competitive Comparison with Loom

Application Scenarios of Loom

• Meeting/Conversation Continuation: When revisiting the same proposal, it automatically recalls the unresolved condition “awaiting Friday data,” enabling discussion to resume seamlessly from the last point.
• Item & Decision Recovery: Instantly query “Where did I leave my keys?” or “Why was this plan shelved?” and retrieve visual and conversational evidence from the original experience.
• Life Experience Review: For questions like “Which places did I visit and what food did I eat over the past year?” it automatically organizes summaries into detailed, evidence-backed retrospectives.
• Personal Preference Learning: Extracts patterns from repeated dining and travel choices to build memory snapshots, enabling highly personalized recommendations.
• Zero-Burden Temporary Reminders: Ideas arising during commutes or tasks discussed in chat don’t need manual note-taking—the system captures them automatically and surfaces them when needed.

Source: AI Tools Hub