Construction AI Safety Tool Gains Long-Term Memory to Prevent Repeat Hazard Warnings

Kaya, a safety copilot for construction sites, uses worker-worn headsets with cameras and microphones alongside CCTV feeds to detect and flag hazards in real time. The system's attention tracker monitors whether workers visually acknowledge hazards, escalating alerts if a danger is ignored for more than four seconds. A core flaw caused the system to forget hazards after five seconds off-screen, repeatedly flagging the same dangers and causing workers to tune out alerts. To fix this, developers integrated Hindsight, an open-source memory layer for AI agents, enabling semantic and time-aware retrieval of past hazard events across all site cameras. A single shared memory bank now allows every device on a site to access what any other camera has previously observed, eliminating redundant warnings.