BlackBox — Persistent Visual Memory

BlackBox is a persistent visual-memory system for cameras and AI agents. Most vision systems describe only what is visible in the current frame. BlackBox instead maintains a memory of objects across time, tracks their state changes, records when they become observed or unobserved, and reasons about what may have happened while they were outside direct view. A user can ask a question such as “Where are my keys?” and BlackBox does not simply guess a location. It combines prior observations, object associations, movement events, and uncertainty to produce a ranked hypothesis with supporting evidence, alternatives, caveats, and a recommendation for what should be inspected next. The system explicitly separates direct observations from inference and treats its confidence values as heuristic weights rather than guaranteed probabilities. The current prototype combines computer vision, object tracking, persistent identities, temporal event detection, a world-state representation, and an interactive web interface. For AMD Developer Hackathon ACT III, BlackBox is being adapted toward an AMD-accelerated inference workflow using ROCm-compatible hardware, with the goal of evaluating how persistent visual-memory pipelines can run efficiently on AMD compute infrastructure. The broader vision is to provide a reusable memory layer for cameras, robots, embodied agents, and other systems that need to reason about a changing physical world over time.