Memories AI – company that creates ‘memories’ for robots and wearables

“AI has worked very well in the digital world. What about the physical world,” Shawn Shen, co-founder of Memories AI, told TechCrunch. “Smart wearable devices and AI-powered robots also need to remember. Ultimately, they need the ability to retain memories.”

Shen believes AI will need to remember what it sees if it wants to “succeed in the physical world”, and that can only be demonstrated through machines equipped with intelligent artificial intelligence. With that goal in mind, Memories AI is using Nvidia’s AI tools to build its own infrastructure.

Earlier this week, Memories AI appeared at Nvidia’s GTC 2026 event as a partner. The company uses the Nvidia Cosmos-Reason 2 inferential visual language model and the Metropolis video search and summarization application to develop its visual memory technology. The cooperation is considered a strategic move to help Shen’s company continue to create “visual memory” technology.

Shen and Ben Zhou, co-founder and Chief Technology Officer of Memories AI, were responsible for building an artificial intelligence system for Meta Ray-Ban glasses. During development, the two thought about how users would actually use the technology in real life if they couldn’t retrieve the video data they were recording. They decided to leave Meta when they saw that the environment was no longer suitable for this goal, then founded the company in 2024 to develop their own technology, headquartered in San Francisco.

The ability to remember and create memories in general is considered a fairly new “field” in the field of intelligent intelligence tools. OpenAI once integrated the feature of remembering old conversations on ChatGPT in 2024 and refined it a year later. Elon Musk’s xAI and Google Gemini also launched similar tools for their AI in the past two years.

According to Shen, these improvements mainly focus on plain-text memory, which is more structured and easier to index. However, it is not useful for physical AI applications, which interact with the world largely through sight and images.

According to Shen, successfully building a “visual memory layer” to help AI remember requires two elements: building the infrastructure needed to embed and index video into a data format that can be stored and retrieved, and collecting the data needed to train the model to do so.

Memories AI launched its Large Visual Memory Model (LVMM) in July 2025 and updated to LVMM 2.0 earlier this year. The model helps users perform a search for a specific moment in less than a second, while empowering them to ask complex questions in natural language or use suggested images to smoothly navigate the video library. In addition, the new version is equipped with highly practical application features such as organizing memories in personal albums or improving security camera analysis and response to real-time situations.

To collect data, the company created Luci, a hardware device that is worn by employees and records video for model training. However, Shen insists they have no intention of becoming a hardware company or selling this device. They built it “just because they weren’t satisfied” with the video cameras available on the market, which focused on high resolution and battery-draining video formats.

Besides Nvidia, the company also cooperated with a number of technology giants, but did not provide details, except for Qualcomm. The company will use this company’s chips for its systems from the end of this year, and plans to introduce a version of LVMM developed specifically for smartphones, virtual reality glasses and personal computers.

Since its launch in 2024, Memories AI has raised $16 million, including an $8 million seed round in July 2025 and an $8 million expansion round a few months later. Investing units include Susa Ventures, Seedcamp, Fusion Fund and Crane Venture Partners.

“In terms of commercialization, Memories AI will focus more on models and infrastructure because we think that eventually the wearables and robotics market will explode,” Shen added. “Of course, maybe not now.”