Memories.ai has evolved into a video data infrastructure company for physical AI, robotics, world models and humanoids. The core product is a datalake: real work recorded at real work sites using the company's own rigs, including a head-mounted stereo pair and a synced wrist view, plus raw IMU data and camera calibration. Every clip passes quality control and is indexed per hand, per action and per task before it joins the lake.
The indexing is what makes the lake searchable by behaviour. Research models like Memories-H0 for hand pose and EgoAction for action segmentation break clips into time-stamped segments, so you can retrieve exactly the moments you need. A query can specify a venue, particular scenes, skill level and a constraint like both hands on the tool, while excluding irrelevant states such as hand idle or walking. Returned moments can then be annotated against your own taxonomy, with dense captions written to the frame per hand.
Delivery is structured rather than raw. Every segment ships as JSON with its hand, label, start and end time and caption, which is the shape robot learning pipelines consume. The company also runs its own capture operation through a network of sites and rigs, offers a wearable 4K first-person camera called LUCI Pin, and has a contributor app where people record tasks and get paid. Teams coming with their own footage can use the indexing and annotation services without contributing to the public lake.