Memories.ai announced on September 24 that its video-language model stack runs natively on Qualcomm Snapdragon platforms. The company presented the approach as a way for its LUCI personal-memory software to capture and retrieve context locally across phones, PCs and, in a technology demonstration, smart glasses.
The engineering idea is to perform visual interpretation when a moment is captured, then use a smaller representation for later searches. That avoids repeating the same heavy processing for every question, according to the company.
Capture and retrieval use different components
The announcement describes OmniCaptioner as the component that turns visual and audio input into compact, time-linked descriptions. OmniRetriever indexes those descriptions and answers retrieval requests against them.
Memories.ai says the models are quantized and compiled for the Hexagon NPU. It describes local processing and a local index as central to the architecture, while allowing connectivity between a person’s devices.
Those are the company’s technical and privacy claims. This report does not independently verify battery use, accuracy or every data path in a finished deployment.
Phone, PC and glasses are at different stages
The company identifies LUCI Mobile on Snapdragon 8 Elite and LUCI Desktop on Snapdragon X2 Elite. It also describes a real-time visual-memory demonstration using Snapdragon AR1 Gen 1 on RayNeo glasses.
The glasses example is explicitly a technology demonstration, not a shipping product. That distinction matters when assessing what a customer can obtain today rather than what the architecture may support later.
The release says the phone can act as a hub for memories from other surfaces. How that behaves for users will depend on device support, permissions and the implementation of cross-device access.
Continuous capture needs understandable controls
A searchable memory of screens or surroundings can contain sensitive information about both the user and other people. Local processing changes where computation happens; it does not remove the need for clear capture, retention and deletion controls.
Useful product checks will include what is recorded, when recording pauses, how exclusions work and which devices can retrieve the resulting records. These are questions for the finished interface, not proof of a defect in the announced system.
For context on the architecture, see our guides to local-first software and NPU hardware.
