Memories.ai brings continuous visual understanding at single-digit billion parameter scale, running on the Qualcomm® Hexagon® NPU inside Snapdragon® platforms.
SAN FRANSISCO, CA, UNITED STATES, September 24, 2026 /EINPresswire.com/ — SAN FRANCISCO, SEPTEMBER 24, 2026 – Memories.ai, the company building the visual memory layer for physical intelligence, today announced at Snapdragon® Summit 2026 that its efficient video language model stack runs natively on Snapdragon® platforms, enabling continuous on-device perception across phone, PC, and smart glasses. The first consumer expression of that stack is LUCI, a personal AI that indexes, remembers, and retrieves a person’s own context on their behalf, with the full perception and retrieval loop executing locally.
The Compute Problem With Continuous Perception
Video understanding has so far been a cloud workload. Frontier vision-language models can run into the tens or hundreds of billions of parameters, and are typically designed to reason over visual inputs that are explicitly provided to them — images or video clips sampled and passed into the model. That paradigm works well for analysing content on demand. But those assumptions start to break down when a device is perceiving continuously, in real time. A wearable or a PC that observes a person’s day produces hours of visual input per day, under a fixed thermal and power budget, on data that most people will not send to a server in the first place. Memories.ai’s approach is to restructure where the compute goes rather than to shrink a cloud model and hope it fits. The stack splits perception into two stages with very different cost profiles.
Architecture: Amortized Perception, Lightweight Retrieval
OmniCaptioner is an efficient video language model that runs at capture time. It converts what a device sees and hears into a compact, structured, temporally grounded description of the moment. This is the only step that touches raw pixels, and it runs once per moment.
OmniRetriever is a multimodal retrieval model that indexes those descriptions into a queryable personal memory, then resolves natural language queries against it at interactive latency.
The design consequence is that heavy visual understanding is amortized at capture rather than paid again at every query. Everything downstream — indexing, retrieval, and answer synthesis — operates on compact representations rather than on video. That keeps every model in the loop at a single-digit billion parameter scale, rather than the tens or hundreds of billions typical of larger models, making the entire pipeline NPU-resident.
Why This Maps Onto Snapdragon®
Models at this scale, quantized and compiled for the Qualcomm® Hexagon® NPU, fit inside the sustained power envelope of a phone, a laptop, and eventually a pair of glasses. That matters for three reasons.
–Continuity: Perception can stay on rather than being invoked. Memory is a function of what was observed, not of what the user remembered to capture.
–Privacy by architecture: The index never leaves the device. The cloud is used only to connect context across a person’s devices when they ask for it, not to store or process the underlying personal content.
–Cost: A continuously running cloud VLM per user is not a viable unit economic. An on-device one is roughly free at the margin.
A full day of continuous memory compresses to a couple of gigabytes on device.
The Use Case: Personal AI
Personal AI is the primary application for on-device perception, and it is the one that has been structurally blocked until now. Most assistants today are stateless. Every session starts from zero context, with no persistent memory of a person’s prior activity, conversations, or environment, because the only memory available to them is what the user types into a chat window.
LUCI closes that gap by giving the model a persistent, queryable memory built from what a person actually sees, hears, and does. LUCI shipped this first on the desktop, indexing on-screen activity in real time. At Snapdragon® Summit, Memories.ai is extending the same architecture across every device a person carries, so a moment captured on one device becomes queryable memory on any of the others, with no centralized cloud index required.
According to Memories.ai founder and CEO Shawn Shen, “The interesting engineering result here is not that we made a model smaller. It is that we moved the expensive part of video understanding to capture time, once, so that everything after it fits on an NPU. The Qualcomm® Hexagon® NPU is what lets that pipeline run continuously and privately, on the device itself. Personal AI is the first place this pays off, but the same stack is what physical intelligence will need generally.”
“At Qualcomm® Technologies, we believe the next generation of AI should be personalized, highly capable, and privacy-first. Memories.ai’s Luci demonstrates how on-device AI can deliver meaningful experiences that understand context and provide intelligent assistance while keeping personal data under a user’s control. By leveraging Snapdragon®‘s AI capabilities across phones, PCs, and smart glasses, Memories.ai is helping unlock a new era of connected, personalized AI experiences that seamlessly enhance user’s everyday experiences,” said Vinesh Sukumar, Vice President, Product Management, Qualcomm® Technologies, Inc.
About Memories.ai
Memories.ai is building the visual memory layer for physical intelligence, technology that helps a model understand the physical world by giving it real, persistent memory rather than raw footage it has to interpret after the fact. LUCI is Memories.ai’s personal AI, built on that same memory engine, giving everyday devices the ability to privately see, remember, and act on a person’s behalf.
Snapdragon® is a trademark or registered trademark of Qualcomm® Incorporated. Snapdragon® is a product of Qualcomm® Technologies, Inc. and/or its subsidiaries. Qualcomm® is a trademark or registered trademark of Qualcomm® Incorporated.
Theodora Lee
Memories.ai
+1 425-615-9692
email us here
Visit us on social media:
LinkedIn
Instagram
YouTube
X
Legal Disclaimer:
EIN Presswire provides this news content “as is” without warranty of any kind. We do not accept any responsibility or liability
for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this
article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
Media gallery
