Memories.ai Drives Physical AI Self-Evolution With Visual Memory
Memories.ai, founded by former Meta Reality Labs researchers Shawn Shen and Ben Zhou, has solidified its position at the forefront of visual memory technology for physical artificial intelligence. Emerging from stealth in July 2025 with an initial $8 million seed round, the Silicon Valley-based firm recently extended its seed financing to $16 million and recruited a chief AI officer from Meta. The company's growth trajectory is marked by significant product milestones and strategic alliances, including the release of its Large Visual Memory Model (LVMM) and the LUCI developer hardware platform showcased at CES in January 2026. Central to Memories.ai's value proposition is its focus on enabling physical AI systems to achieve recursive self-evolution through visual memory, a capability distinct from the commoditized personalization features dominating the current AI landscape. While text-based memory has become standardized, visual data remains structurally complex. Memories.ai addresses this by operating at the codec level, compressing video, converting it to vectors, and generating structured descriptions. The company has built a proprietary video data lake indexing over one million hours of footage, boasting a memory capacity 100 times greater than comparable solutions. This infrastructure supports a self-improvement loop where models deployed in new environments automatically identify failure patterns from visual records, triggering targeted post-training without reliance on manual data annotation. Founder Shen Junxiao projects that this process can boost model accuracy from an initial 70% to nearly 90% with each iteration delivering marginal yet cumulative gains. Strategic partnerships underscore the technology's viability and scalability. In November 2025, Qualcomm agreed to natively support LVMM 2.0 on its processors starting in 2026, enabling edge-side compression and indexing to reduce latency and cloud costs. By March 2026, NVIDIA integrated Memories.ai's technology into its Cosmos-Reason 2 and Metropolis platforms, validating the approach for robotics and spatial computing applications. The company also highlighted recent research advancements, including MARC for token compression, O-MARC for multimodal inference efficiency, and SpatialMem for 3D spatial indexing. Memories.ai has defined a clear business ecosystem centered on a data flywheel. Enterprise clients, such as factories and retail chains, utilize the platform for operational monitoring and standard procedure checks, allowing real-world video data to accumulate organically in the data lake. This corpus is then shared with advanced robotics laboratories developing vision-language-action models. The resulting AI systems are deployed back into enterprise environments, closing the loop between data collection, model refinement, and physical application. Shen emphasizes a strict service orientation; the company provides data assessment and model post-training without competing in robot manufacturing or foundational model development. The firm maintains a lean organizational structure prioritizing high-agency talent and vision alignment. Shen advocates for a scaling strategy driven by high-quality real-world data over simulation, arguing that physical AI must learn from actual deployment conditions. As foundational model capabilities continue to expand exponentially, Memories.ai positions visual memory not as an end product but as the critical enabling layer for autonomous systems capable of independent adaptation and long-horizon task execution in unstructured physical environments.
