Watch Skill: An Open-Source Local Video Memory Layer for AI Agents

Fearless-Role-2707 · reddit · 2026-08-10

A developer has open-sourced Watch Skill, a local-first video memory layer project designed to give AI agents efficient recall of long video content.

Core Architecture: Transforms videos into scene-aware frames with transcripts/OCR to build a local index. Agents can later retrieve relevant timestamped clips without reprocessing the entire video.
Fully Local Pipeline: Transcription, OCR, retrieval, and embedding searches run entirely locally. Visual reasoning integrates seamlessly with Ollama, making cloud VLMs completely optional.
Agent Workflow Integration: In automated coding tasks, agents can use recordings of browser or desktop sessions as execution traces, pinpointing exact failure points and enabling a record-inspect-fix loop.
Built-in Benchmarking: Includes a visual model benchmark that evaluates configured models against identical frames for latency, accuracy, and token cost.

More from coding & agent