What problem does it solve? Standard video reading tools sample too few frames to meaningfully cover videos longer than 30 minutes, so questions about multi-hour footage go unanswered. This Skill builds a persistent hierarchical graph memory (Root, SuperEvent, MacroEvent, Subgraph) for long videos, letting you locate and inspect any segment through semantic search instead of re-scanning the whole file. ## Core Features & Use Cases - Hierarchical Memory Construction: Segments long videos via HLS frame-diff scene detection, extracts entities, events, OCR text, and relations per segment with a vision-language model, and aggregates them into a 4-level tree stored in <video_path>.memory/. - Semantic Retrieval Tools: Query memory with search_nodes, search_ocr_text, search_asr_text, search_by_time, and enumerate_events, then drill down with get_subgraph for full detail on a located segment. - Parallel Build Pipeline: Splits long videos into chunks processed concurrently with checkpointing, resume support, ASR transcription, and hybrid dense-plus-BM25 embedding indexes. - Use Case: Given a 3-hour lecture recording, build its memory once, then ask "What did the speaker say about gradient descent?" to locate the exact segment and inspect frames with read_video over a narrow time range. ## Quick Start Ask the agent to build memory for your long video with build_memory.sh and then answer questions about its content using the video-memory tools.
