Memories.ai and NVIDIA Partner to Grant Robots and Wearables Permanent Visual Memory

The Future of Visual Recall

Memories.ai, an AI startup founded by former Meta researchers, has announced a strategic collaboration with NVIDIA to bring persistent visual memory to hardware. This partnership pairs the Memories.ai Large Visual Memory Model (LVMM) with NVIDIA Cosmos Reason 2 and the NVIDIA Metropolis platform. By integrating these technologies, robots and wearables can now see, remember, and learn from the physical world in real time.

Most current AI systems are stateless, meaning they process information in the moment but fail to retain long-term context. This AI amnesia prevents smart glasses from remembering last week's events or robots from learning a home's layout permanently. Memories.ai addresses this bottleneck by turning continuous video into structured, searchable memory atoms that span long durations.

The collaboration utilizes NVIDIA RTX GPUs for personal computers and NVIDIA Jetson Orin for mobile devices. These hardware platforms allow the LVMM to run efficiently at the edge. On a laptop, this enables AI agents to draft emails based on past meetings it saw or retrieve files from previous workflows based on visual history.

Smart glasses powered by this technology provide proactive assistance, such as reminding users of names or surfacing documents before a meeting starts. In robotics, the memory model allows machines to remember spatial layouts and obstacle patterns across different sessions. This ensures that every interaction with the physical world makes the device smarter rather than forcing it to restart from scratch.

To accelerate adoption, Memories.ai introduced Project LUCI, an open-source reference platform for AI wearables. Weighing under 45 grams, the device serves as a testbed for hardware makers to build context-aware systems. The company aims for LUCI to become a standardized infrastructure for visual recall, similar to how Android serves mobile phones.

Founded in 2024, Memories.ai has already indexed over one million hours of video to train its foundational models. The startup recently secured a total of 16 million dollars in funding from investors like Susa Ventures and Samsung Next. This financial backing supports their mission to make the physical world as searchable as the digital one through advanced memory graphs.