Not long after working on Meta's Ray-Ban smart glasses at Reality Labs, two engineers began asking an uncomfortable question: What good is computer vision if the computer can't remember what it saw yesterday?
That line of thinking has now pulled in $16 million. Memories.ai, the San Francisco startup founded by Dr. Shawn Shen and Enmin "Ben" Zhou, closed an $8 million seed round last July—Susa Ventures led, with Samsung Next, Seedcamp, Fusion Fund, Crane Venture Partners, and Creator Ventures joining in. Eight months later, in March, another $8 million landed, this time an extension with much of the original cast returning.
The pitch is deceptively simple: build what the founders call a Large Visual Memory Model, an infrastructure layer that lets AI systems retain and recall visual information across millions of hours of footage. Think of it as giving machines something closer to episodic memory—the ability to not just see, but to remember seeing.
Right now, AI can analyze video. It can tag objects, track movement, even generate descriptions. What it generally can't do is sift through months of footage and answer a question like, "Show me every time someone parked a white van near the loading dock after 9 p.m." At least not without someone laboriously training a model on that specific task, or a security analyst burning hours scrubbing through tape.
Memories.ai's LVMM compresses, indexes, and retrieves visual data across timelines that stretch far beyond what current systems handle comfortably. The company speaks in terms of weeks, months, years of footage—stored, searchable, usable.
From Menlo Park to Market
Shen and Zhou both did stints at Meta Reality Labs, the division responsible for the company's mixed-reality hardware and the AI that powers it. Their work on the Ray-Ban Meta glasses gave them firsthand insight into how visual AI performs in the real world and where it still stumbles.
The founding timeline is a little murky. TechCrunch reported the company launched in 2024; LinkedIn suggests 2025, though TechCrunch appears more reliable. Either way, by mid-2025 the startup had paying customers and a clear thesis: that AI's inability to maintain long-term visual context was a bigger bottleneck than most investors realized.
Security Footage and Brand Mentions
Early adopters cluster in two camps. Physical security firms use the platform to query archives of CCTV footage, flagging anomalies or unusual patterns without needing a human to sit through hours of tape. Marketing intelligence teams, meanwhile, deploy it to parse social video—tracking brand mentions, analyzing product placements, identifying emerging trends across thousands of hours of creator content.
By July 2025, according to Samsung Next, the company had indexed 1 million hours of video. Two months later, in September, Memories.ai claimed that figure had grown past 10 million hours. The ramp-up suggests early traction, though the company hasn't disclosed revenue or customer count.
The roadmap, however, extends well past cloud-based video analytics.
Edge Cases
In November 2025, Memories.ai announced it was working with Qualcomm to bring its LVMM 2.0 on-device—targeting wearables, robots, and other edge deployments where sending video to the cloud isn't practical or fast enough. A few months later, in March, the company began using NVIDIA's Cosmos-Reason 2 model and Metropolis architecture, and announced an integration with Ring's network of what the company claims is more than 100 million cameras.
That's a lot of surface area for a startup that, according to LinkedIn, employs somewhere between 11 and 50 people.
Perhaps the most telling move came at Microsoft Build this past June, where Memories.ai demoed Project LUCI—an on-device visual memory system running across PCs, wearables, and IoT hardware using Windows ML and Qualcomm's X2 Elite processors. The shift from cloud to edge suggests the company sees its future less in video surveillance dashboards and more in ambient computing: AI that quietly remembers what you've seen, no matter where you saw it.
It's an ambitious pivot. And one that raises familiar questions about privacy, consent, and just how much visual data people are comfortable letting an AI system retain on their behalf. The founders haven't said much publicly about those guardrails yet—though given the regulatory scrutiny around wearable AI and smart glasses, they'll need to.
For now, the funding gives them runway to chase use cases the company has flagged publicly: media production, sports analytics, and the emerging wearables and robotics markets. Whether visual memory becomes the infrastructure layer the founders envision, or just another niche tool in an already crowded computer vision market, likely depends on how well they can convince customers—and regulators—that teaching AI to remember is worth the tradeoffs.
