Former Meta Researchers Secure $8 Million for Memories.ai to Build AI That Understands Long Videos

Memory Revolution: How Memories.ai's $8M Seed Round Could Transform the Future of Video Intelligence

In a world drowning in video data, a small team of ex-Meta researchers believes they've solved one of AI's most stubborn limitations: the inability to truly understand lengthy video content. Their solution could reshape industries from security to entertainment, while creating new investment frontiers in AI infrastructure.

The Forgotten Dimension of Machine Vision

Memories.ai, founded by former Meta Reality Labs researchers Dr. Shawn Shen and Enmin Zhou, emerged from stealth on July 24, 2025, with an $8 million seed funding round. The round was led by Susa Ventures with participation from Samsung Next, Fusion Fund, Crane Venture Partners, Seedcamp, and Creator Ventures.

The startup has developed what they call a Large Visual Memory Model (LVMM), designed to overcome limitations in current AI systems that struggle to process more than a few hours of video footage. According to the company, their technology can scale to analyze up to 10 million hours of video - far beyond existing capabilities in the industry.

The funding round—twice what the company originally targeted—was led by Susa Ventures with participation from Samsung Next, Fusion Fund, Crane Venture Partners, Seedcamp, and Creator Ventures.

Digital Recall: The Billion-Dollar Problem Hiding in Plain Sight

The limitation Memories.ai addresses represents a massive inefficiency across multiple industries. While text-based AI has seen models capable of processing hundreds of thousands of tokens, video analysis typically fails after just minutes of content.

"Companies are sitting on petabytes of video they can't effectively search or analyze," notes a senior technology analyst at a major investment bank. "Security teams watch endless footage manually. Media companies can't find scenes in their own archives. Marketing departments struggle to extract insights from campaign videos."

This inefficiency translates to a substantial market opportunity. The global video analytics market, valued at approximately $12.33 billion in 2024, is projected to reach nearly $94.56 billion by 2034, growing at a 22.6% CAGR over that period.

Computational Memory that Scales with Reality

What distinguishes Memories.ai's approach is both scale and architecture. The LVMM first ingests and compresses raw video into a structured memory layer, then builds contextual relationships between visual elements across arbitrary timeframes.

"Traditional models analyze video frame-by-frame or in short clips, losing all context between segments," explains Zhou. "Our system mimics human memory—retaining important information while filtering out noise, creating connections between related events, and allowing natural language retrieval of specific moments."

The technology can reportedly scale to process up to 10 million hours of footage—orders of magnitude beyond current capabilities. More impressively, much of this processing can happen directly on user devices rather than exclusively in the cloud.

From Security Cameras to Hollywood Archives

Early applications focus on sectors drowning in video data:

In security and surveillance, the system can instantly surface relevant footage across vast archives—potentially transforming incident investigation from days to minutes. For media companies, it promises to make massive content libraries instantly searchable by scene, prop, character, or action.

Marketing teams can analyze trends across thousands of social media videos, while robotics companies see potential for machines that learn continuously from visual experiences.

Perhaps most intriguing for consumer applications, the technology could allow users to search their personal video collections with natural language queries like "find the video of my daughter's first steps" or "show me all beach sunsets from our vacations."

The Race for Video Memory Dominance

Memories.ai isn't alone in recognizing this opportunity. TwelveLabs has raised $80 million across multiple rounds, including participation from NEA, NVIDIA NVentures, Databricks, Snowflake, and others. Their technology, however, is currently optimized for videos up to only 60 minutes in length.

Other competitors include mem0 (YC-backed, focused primarily on text memory), Letta ($10M seed from Felicis), and offerings from tech giants like Google's Video Intelligence API and Amazon Rekognition.

"What sets Memories.ai apart is both the scale they're targeting and their on-device capabilities," observes a venture partner specializing in AI investments. "If they can deliver even half of what they're promising with acceptable performance, they'll leapfrog today's incumbents."

The Path Forward: Opportunities and Challenges

Despite the promising technology and substantial market, Memories.ai faces significant hurdles. The company must prove its system can maintain accuracy and performance at scale in real-world deployments—not just controlled demos.

Building an enterprise sales operation represents another challenge for the research-heavy team, which currently numbers just 15 employees. Managing the computational costs of processing millions of video hours while remaining price-competitive presents yet another obstacle.

"The window for creating a defensible moat is short," cautions a partner at a leading technology investment firm. "Microsoft's Copilot Memory launches next month, and both Google and Amazon are rapidly expanding their video capabilities."

Investment Horizon: Reading the Signals

For investors eyeing the video intelligence sector, several indicators may signal Memories.ai's trajectory:

Conversion metrics: Watch for announcements of pilot programs converting to six-figure annual contracts, particularly in security and media verticals where the pain point is most acute.

Edge deployment success: The ability to perform substantial analysis on-device represents a potential competitive advantage—especially as privacy regulations tighten globally.

Unit economics: Processing efficiency will prove crucial; the computational cost per hour of video must decrease substantially to maintain healthy margins at scale.

Strategic partnerships: Look for potential collaborations with device manufacturers (leveraging on-device capabilities) or security system integrators that could accelerate enterprise adoption.

Market observers suggest three potential outcomes: strategic acquisition by a larger player (Adobe, NVIDIA, or AWS being likely candidates), growth toward an independent public offering, or integration into edge hardware as a "memory SDK."

While the technology appears promising, institutional investors should conduct thorough technical validation before committing substantial capital. Testing the system against a cold dataset of 40,000+ hours of footage and comparing precision/recall metrics against competitors would provide valuable insight into whether Memories.ai can deliver on its ambitious claims.

As enterprises and consumers alike struggle with ever-increasing volumes of video data, the race to build machines that truly understand and remember what they see promises to reshape both technology markets and user experiences in the years ahead.