Memories AI: Building Visual Memory for Wearables and Robotics
▼ Summary
– Shawn Shen’s company Memories.ai is building AI infrastructure for visual memory, which he believes is essential for wearables and robotics to succeed in the physical world.
– The company is developing this technology using Nvidia’s AI tools and has raised $16 million in funding since its 2024 launch.
– The founders got the idea while working on Meta’s Ray-Ban glasses, realizing a need for AI that can recall recorded video data.
– Their solution involves a Large Visual Memory Model (LVMM) and custom hardware for data collection, focusing on the underlying model and infrastructure rather than selling devices.
– While current AI memory advances are largely text-based, Memories.ai is targeting the future commercialization of visual memory for the wearables and robotics markets.
The future of artificial intelligence in our physical environment hinges on its ability to remember what it sees. Memories.ai is building the essential visual memory infrastructure for wearables and robotics, a foundational layer that allows AI to recall and utilize video data. The company recently announced a strategic collaboration with Nvidia, leveraging their advanced AI tools to accelerate this development.
Founder Shawn Shen and CTO Ben Zhou conceived the idea while developing the AI system for Meta’s Ray-Ban smart glasses. They realized a critical gap: users could record video, but the AI lacked the capability to remember and reference that visual information later. After searching unsuccessfully for an existing solution, they decided to leave Meta and build it themselves. Shen argues that while AI excels in the digital realm, success in the physical world demands a new kind of intelligence. For AI wearables and robots to be truly useful, they must possess visual memories, not just text-based recall.
While memory features are becoming more common in large language models, these advancements have primarily focused on text. Text is structured and easy to index, but it falls short for devices that interact with the world primarily through sight. Shen’s vision requires a system that can process, embed, and retrieve unstructured visual data efficiently.
To bring this vision to life, the company tackled two major challenges: creating the infrastructure to index video into a storable format and gathering the necessary training data. They launched their large visual memory model (LVMM) in July 2025, which Shen describes as a compact counterpart to other multimodal indexing models. For data collection, they developed a custom hardware device called LUCI, worn by dedicated data collectors to record training videos. The company emphasizes it is not becoming a hardware manufacturer; they built LUCI because consumer recorders, optimized for high-definition playback, were unsuitable for their AI training needs.
Memories.ai has already secured $16 million in funding and has released a second-generation LVMM. A new partnership with Qualcomm will see the model running on their processors later this year. Although Shen confirms work with major wearable companies, their names remain confidential. The immediate commercial focus is on perfecting the model and infrastructure, anticipating that the mass market for advanced AI wearables and robotics is on the horizon, even if it hasn’t fully arrived today.
(Source: TechCrunch)
