Memories.ai moves continuous visual AI onto Snapdragon chips

Memories.ai said at Snapdragon Summit 2026 that its video language stack now runs natively on Snapdragon platforms, enabling continuous on-device perception across phones, PCs and smart glasses. The move is meant to keep personal context local while powering its LUCI personal AI and a broader push toward physical intelligence.

Why it matters: - Memories.ai is trying to make continuous visual understanding practical on consumer devices without relying on the cloud. - The approach could improve privacy, lower inference costs and keep personal memory searchable across a person's devices. - The company is positioning on-device perception as a core building block for personal AI and physical intelligence.

What happened: - Memories.ai announced at Snapdragon Summit 2026 that its efficient video language model stack runs natively on Snapdragon platforms. - The stack runs entirely on the Qualcomm Hexagon NPU inside Snapdragon platforms. - The first consumer product built on the stack is LUCI, a personal AI that indexes, remembers and retrieves a person's context locally. - The announcement covers phones, PCs and smart glasses.

The details: - OmniCaptioner is the capture-time model. - OmniCaptioner converts what a device sees and hears into a compact, structured, temporally grounded description of the moment. - OmniCaptioner is the only step that touches raw pixels. - OmniRetriever indexes those descriptions into a queryable personal memory. - OmniRetriever resolves natural language queries at interactive latency. - The system shifts heavy visual understanding to capture time instead of repeating it at every query. - Downstream indexing, retrieval and answer synthesis operate on compact representations rather than video. - The models stay in the hundreds of millions of parameters rather than the billions. - That scale is what keeps the pipeline NPU resident. - Models at this scale can be quantized and compiled for the Qualcomm Hexagon NPU. - Memories.ai said a full day of continuous memory compresses to a couple of gigabytes on device.

Between the lines: - Memories.ai is betting that persistent memory is the missing feature in today’s assistants, which usually start every session from zero. - The architecture is designed to avoid a centralized cloud index, which strengthens the company’s privacy pitch. - The company is also arguing that cloud-based continuous vision is not economical for always-on personal use. - The Qualcomm partnership gives Memories.ai a credible hardware path for sustained on-device AI across multiple form factors. - Shawn Shen framed the engineering shift as moving expensive video understanding to capture time so the rest can run on an NPU. - Qualcomm Vice President of Product Management Vinesh Sukumar said the demo shows how on-device AI can keep personal data under user control while delivering context-aware assistance.

What's next: - Memories.ai said LUCI shipped first on desktop, indexing on-screen activity in real time. - The company is extending the same architecture across devices so a moment captured on one device can be queried on another. - The cloud may still connect context across a person’s devices when needed, but Memories.ai does not plan to use it as the primary storage layer for personal content. - The broader goal is to scale the same memory stack into future physical intelligence products.

The bottom line: - Memories.ai is turning continuous visual memory from a cloud problem into an on-device Snapdragon use case, with privacy and cost as the main selling points.