Startups & Business News

Samsung-Backed Memories.ai Redefines AI Video Search and Privacy

Aiko Tan
July 29, 2025

KEY POINTS

Memories.ai, a San Francisco-based AI startup founded by former Meta researchers, raised $8 million in seed funding backed by Samsung Next and Susa Ventures to push the boundaries of video intelligence.
Its signature innovation, the Large Visual Memory Model (LVMM), is designed to process, compress, and make searchable up to 10 million hours of video—a massive leap over the 1-2 hour limitations facing most current AI models.
The platform enables natural-language search across long-form video, serving use cases from security (instant video review) to marketing (brand trend analysis).
On-device processing is a standout, addressing privacy concerns and reducing dependency on cloud-based storage—key for enterprise and consumer security settings.

Breaking the Hour Barrier: Why Memories.ai Stands Out

For years, even the best artificial intelligence (AI) could barely “remember” an hour or two of video before losing context—limiting usefulness for sectors like security, marketing, and digital media. Memories.ai, helmed by ex-Meta Reality Labs researchers Dr. Shawn Shen and Ben Zhou, aimed to change that. Recent funding—$8 million led by Susa Ventures and Samsung Next—shows that investors believe this “memory layer for machines” is more than hype.

The Large Visual Memory Model (LVMM): Unpacking the Tech

The company’s Large Visual Memory Model (LVMM) has one mission: give machines true, persistent “visual memory.” Instead of just summarizing one video at a time, LVMM indexes, compresses, and makes searchable millions of hours of video, enabling sophisticated natural language queries and instant recall.

From a technical standpoint, Memories.ai mirrors how human memory works:

Cues trigger recall.
A query model retrieves relevant segments.
A selection model extracts details and integrates them for answers.
The system self-corrects through “reflection,” ensuring accuracy and context are maintained over long video timelines.

The result: users can instantly ask the AI questions like, “Show me every breach at Entrance A last year,” or “Which influencer featured our product in TikTok videos this quarter?”, spanning video archives that used to be functionally invisible.

On-Device AI: Samsung’s Privacy Play

Samsung Next’s involvement isn’t just about the power of big data. On-device video processing means companies (or even consumers) don’t have to upload sensitive footage to the cloud to benefit from AI search and analysis—solving a major privacy hurdle for those wary of surveillance technology. Edge AI like this has obvious Samsung Galaxy, smart home, and security product implications for the future.

Real-World Applications: From Security to Brand Intelligence

Security: Instantly scan months of surveillance for specific people, objects, or incidents. No more manual scrubbing or “lost” evidence.
Marketing: Analyze trending visual themes, brand visibility, and influencer activity across social video at scale.
Entertainment & Media: Navigate massive content archives, instantly surfacing scenes, shots, or dialogue across years’ worth of footage.
Potential Future Roles: LVMM could power “memories” for personal assistants, guide training for robots and autonomous vehicles, and more.

Competitive Pressure and the Road Ahead

Memories.ai isn’t alone in the space:

TwelveLabs and Google lead on the deep learning side.
Competitors like mem0 and Letta offer memory layers, but with limited video capabilities.

Still, Memories.ai’s horizontal, model-neutral platform—combined with private, on-device analysis—is a key differentiator.

Looking Forward

Armed with fresh funding, the 15-person Memories.ai team is expanding fast, aiming to make video memory and search as fundamental—and accessible—as language models are today. Whether it’s security, analytics, or powering the AI agents of tomorrow, persistent visual memory could become an essential part of the AI toolkit—even outside of news headlines.