Memories.ai

Memories.ai builds a video data infrastructure for physical AI, offering a scalable Datalake that grows with your data. The system records video from wearables and rigs at real work sites, with synchronized views from head-mounted and wrist cameras. Clips are indexed by hand and action and arrive ready for training pipelines. The company positions this as an infrastructure product rather than a vendor offering, aiming to supply continuous data for model improvement.

Memories.ai describes its workflow as capturing raw video through its own rigs, curating usable content by querying and annotating only relevant moments, and delivering annotations in a structured format for training. Each hand’s actions are time-stamped, and captions are created for every segment against a user-defined taxonomy. The end result is a set of ready-to-use training data that can be integrated into existing pipelines.

The main focus is to provide dense captions, per-segment annotations, and JSON-delivered metadata aligned to video moments. The platform emphasizes action segmentation, frame-accurate captions, and a delivery format that includes hand, label, start and end times, and text. This metadata is designed to fuel robot learning and AI systems that rely on precise, frame-true annotations.

Memories.ai emphasizes enterprise readiness and real-work data collection, noting that it captures with wearable rigs, stores in a Datalake, and offers tools to search, annotate, and deliver data for model training. The platform is presented as capable of handling large-scale, diverse video data relevant to physical AI applications.

Use cases

• Provide training data for robot learning pipelines with per-hand segmentation and frame-accurate captions.
• Use the Video Datalake to support world models and long-horizon dexterous hand skills in robotics.