Memories.ai and Nvidia give robots visual memory

If you wear smart glasses or watch a robotic vacuum roam your living room, you run into the same awkward reality: physical devices have cameras everywhere, but they suffer from gold-fish-level visual amnesia.

AI assistants can now easily recall what you texted them three weeks ago. But ask your smart glasses where you left your car keys ten minutes ago, and they draw a complete blank. The reason is simple: almost all modern AI memory infrastructure was built purely for text.

On March 16 at Nvidia's GTC conference, startup Memories.ai announced a partnership with the chip giant. Founded by former Meta engineers, the company is plugging this gap by building a dedicated visual memory layer designed specifically for wearables and robotics.

Born out of Meta's smart glasses team

Co-founders Shawn Shen and Ben Zhou came up with the idea while working on the AI system behind Meta’s Ray-Ban glasses. They realized pretty quickly that shooting endless first-person footage is pointless if users can never actually search or retrieve what was recorded.

When they couldn't find anyone building visual-first recall tools, they left Meta in 2024 to build it themselves. While OpenAI, xAI, and Google have rolled out text memory over the past two years, Shen argues that text is neat and easy to index, while physical AI has to navigate the messy visual world.

The startup has raised $16 million to date, including an $8 million seed round in July 2025 and an $8 million extension led by Susa Ventures, with participation from Seedcamp, Fusion Fund, and Crane Venture Partners.

Teaming up with Nvidia and Qualcomm

Building visual memory basically boils down to two things: an efficient way to embed and index video streams, and enough real-world footage to train the system.

Under the new partnership, Memories.ai is tapping Nvidia's Cosmos-Reason 2 vision-language reasoning model along with the Nvidia Metropolis stack for scalable video search and summarization.

On the model front, the company released its first large visual memory model (LVMM) in July 2025, which Shen compares to a lightweight version of Google's Gemini Embedding 2. They have now rolled out their second-generation LVMM, and they already signed a deal with Qualcomm to run it directly on Qualcomm processors starting later this year.

Interestingly, they even built custom hardware called LUCI just for internal data collectors to record training video. Shen made it clear they aren't getting into the consumer gadget business; off-the-shelf action cameras were simply too battery-hungry and high-res for efficient ML collection.

Why it matters

Giving AI eyes with a working memory tackles one of the biggest friction points in physical and embodied AI today.

Right now, edge devices are trapped in a weird middle ground: powerful vision models can identify objects in real time, but storing and recalling hours of raw footage crushes device battery life and cloud storage. By positioning itself as a dedicated indexing layer between sensors and reasoning engines, Memories.ai is carving out a smart infrastructure play without taking on the crushing capex of building foundation models or selling hardware.

For end users, this is the exact missing link needed to make AI glasses genuinely helpful for finding lost items or recalling spoken conversations. But don't expect it in consumer hands tomorrow. Shen admits the big wave for wearables and robotics is coming, but "it's probably just not now."

We think their decision to partner with both Nvidia on the cloud side and Qualcomm on the edge side is strategically sharp. The real hurdle will be convincing device makers that this indexing can run without draining tiny batteries or triggering massive privacy backlashes. Still, while the rest of the industry obsesses over chat context windows, building the visual recall layer for the real world is a bet worth making.