Imported from PixelML/av (skills/agentic-video-memory/SKILL.md). Install upstream with npx skills add PixelML/av --skill agentic-video-memory. Copyright stays with the author.

Agentic Video Memory

You are setting up av — a CLI that turns video files into searchable, queryable memory. Index once, search many times. What jq is for JSON, but for video.

Quick Start

Run these three commands:

pip install pixelml-av
av config setup
av ingest <video-file-or-url>

After ingestion, you can search and ask questions:

av search "what was discussed about pricing"
av ask "what were the key decisions?"

Requirements

• Python 3.11+
• ffmpeg — must be installed and on PATH (brew install ffmpeg / apt install ffmpeg)
• API key — one of: OpenAI, Anthropic (Claude), or Google (Gemini)

Configuration

Interactive Setup

Run av config setup to launch the interactive wizard. It will ask you to:

1. Choose a provider (OpenAI, Anthropic, Gemini)
2. Enter your API key
3. Validate the key works

Configuration is saved to ~/.config/av/config.json.

Environment Variable Overrides

You can override any config value with environment variables. These take priority over config.json.

Provider Capabilities

When a capability is unavailable (e.g., Anthropic has no transcription), the pipeline skips that stage and warns. Use AV_OPENAI_API_KEY as a transcription fallback for non-OpenAI providers.

Commands

av ingest <path>

Ingest a video file, directory, or YouTube URL into the index.

# Single file
av ingest meeting.mp4

# With frame captions (vision model describes video frames)
av ingest meeting.mp4 --captions

# Dense visual captioning timeline
av ingest meeting.mp4 --dense-vision

# YouTube URL
av ingest "https://youtu.be/dQw4w9WgXcQ"

# Batch — ingest all videos in a directory
av ingest /path/to/videos/

# Skip embedding generation (transcript only)
av ingest meeting.mp4 --no-embed

# Re-ingest even if file hash matches
av ingest meeting.mp4 --force

# Preview what would happen without executing
av ingest meeting.mp4 --dry-run

Flags:

Output (stdout, JSON):

{
  "status": "complete",
  "video_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "filename": "meeting.mp4",
  "duration_sec": 3600.0,
  "artifacts_count": 847,
  "elapsed_sec": 45.2
}

On partial failure:

{
  "status": "complete_with_warnings",
  "video_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "filename": "meeting.mp4",
  "duration_sec": 3600.0,
  "artifacts_count": 320,
  "elapsed_sec": 30.1,
  "warnings": ["Transcription skipped: provider does not support whisper"]
}

av search <query>

Full-text + semantic search across all indexed videos.

av search "what was discussed about pricing"
av search "action items" --limit 5
av search "deployment plan" --video-id a1b2c3d4

Flags:

Output (stdout, JSON):

{
  "query": "what was discussed about pricing",
  "results": [
    {
      "rank": 1,
      "score": 0.87,
      "video_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
      "filename": "meeting.mp4",
      "timestamp_sec": 1455.0,
      "timestamp_formatted": "00:24:15",
      "source_type": "transcript",
      "text": "We agreed on the $49/mo tier for the starter plan...",
      "artifact_id": "f1e2d3c4-b5a6-7890-abcd-ef1234567890"
    }
  ],
  "total_results": 5,
  "search_time_ms": 12
}

av ask <question>

RAG question-answering over your indexed videos. Returns an answer with timestamped citations.

av ask "what were the key decisions?"
av ask "summarize the meeting" --video-id a1b2c3d4
av ask "what did they say about the budget?" --top-k 20

Flags:

Output (stdout, JSON):

{
  "answer": "Three key decisions were made in the meeting: 1) Launch the starter plan at $49/mo, 2) Push the release to Q2, 3) Hire two more engineers for the platform team.",
  "citations": [
    {
      "video_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
      "start_sec": 1455.0,
      "source_type": "transcript",
      "text": "We agreed on the $49/mo tier for the starter plan...",
      "score": 0.91
    }
  ],
  "confidence": 0.85
}

av list

List all indexed videos.

av list

Output (stdout, JSON):

{
  "videos": [
    {
      "video_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
      "filename": "meeting.mp4",
      "duration_formatted": "01:00:00",
      "status": "complete",
      "artifacts_count": 847
    }
  ],
  "total": 1
}

av info <video_id>

Show detailed metadata for a specific video.

av info a1b2c3d4-e5f6-7890-abcd-ef1234567890

Output (stdout, JSON):

{
  "video_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "filename": "meeting.mp4",
  "file_path": "/Users/you/Videos/meeting.mp4",
  "duration_sec": 3600.0,
  "duration_formatted": "01:00:00",
  "resolution": "1920x1080",
  "status": "complete",
  "artifacts": {
    "transcript": 320,
    "caption": 42
  },
  "ingested_at": "2026-02-14T10:30:00Z"
}

av transcript <video_id>

Output the transcript for a video.

av transcript a1b2c3d4 --format vtt
av transcript a1b2c3d4 --format srt
av transcript a1b2c3d4 --format text

Flags:

Output (stdout, text — not JSON):

WEBVTT

00:00:00.000 --> 00:00:05.500
Welcome to the Q1 planning meeting

00:00:05.500 --> 00:00:10.000
Today we'll discuss the product roadmap

av export

Export all video memory as JSONL or subtitle formats.

av export
av export --format jsonl
av export --format srt --video-id a1b2c3d4

Flags:

Output (stdout, JSONL — one JSON object per line):

{"video_id": "a1b2c3d4", "filename": "meeting.mp4", "type": "transcript", "start_sec": 0.0, "end_sec": 5.5, "timestamp_formatted": "00:00:00", "text": "Welcome to the Q1 planning meeting"}
{"video_id": "a1b2c3d4", "filename": "meeting.mp4", "type": "transcript", "start_sec": 5.5, "end_sec": 10.0, "timestamp_formatted": "00:00:05", "text": "Today we'll discuss the product roadmap"}

av open <video_id>

Open a video file at a specific timestamp.

av open a1b2c3d4
av open a1b2c3d4 --at 90.5

Flags:

av config setup

Interactive wizard to configure your AI provider and API key. No flags — fully interactive.

av config show

Show current configuration.

Output (stdout, JSON):

{
  "provider": "openai",
  "api_base_url": "https://api.openai.com/v1",
  "api_key": "***",
  "transcribe_model": "whisper-1",
  "vision_model": "gpt-4-1",
  "embed_model": "text-embedding-3-small",
  "chat_model": "gpt-4-1",
  "db_path": "~/.config/av/av.db"
}

av config path

Show the path to the database file.

Output (stdout, plain text):

/Users/you/.config/av/av.db

av version

Print version info.

Output (stdout, JSON):

{"version": "0.1.0", "package": "pixelml-av"}

Agent Integration Patterns

JSON stdout / stderr separation

All data-producing commands output valid JSON to stdout. Progress messages, warnings, and errors go to stderr. This means you can safely parse stdout without filtering out log noise.

# Capture just the JSON output
result=$(av search "pricing" 2>/dev/null)

# Or pipe directly to jq
av search "pricing" 2>/dev/null | jq '.results[0].text'

Batch directory ingestion

Point av ingest at a directory to ingest all video files:

av ingest /path/to/recordings/

Supported video extensions: .mp4, .mkv, .avi, .mov, .webm, .flv, .wmv, .m4v, .mpg, .mpeg, .3gp, .ts

Chaining search and ask

Use search to find relevant videos, then ask questions about specific ones:

# Find which video mentions pricing
av search "pricing discussion" 2>/dev/null | jq -r '.results[0].video_id'

# Ask a question about that specific video
av ask "what was the final price agreed?" --video-id <video_id_from_above>

Piping to other tools

# Export transcript and pipe to another tool
av transcript <video_id> --format text | wc -w

# Export all data as JSONL for processing
av export --format jsonl > all_videos.jsonl

Enterprise

Pixel ML offers managed video intelligence infrastructure for teams that need:

• Custom model hosting and fine-tuning
• SLA guarantees and priority support
• On-premise deployment
• High-volume ingestion pipelines

Contact hello@pixelml.com for enterprise pricing and setup.

Troubleshooting

ffmpeg: command not found

Install ffmpeg:

• macOS: brew install ffmpeg
• Ubuntu/Debian: sudo apt install ffmpeg
• Windows: Download from https://ffmpeg.org/download.html

av: command not found after pip install

Make sure your Python scripts directory is on PATH. Try:

python -m av --help

Or install with pipx:

pipx install pixelml-av

API key errors / authentication failures

Run av config setup again to re-enter your API key. Check that:

• The key is valid and not expired
• The key has access to the models listed in the provider table above
• You're using the correct provider

Transcription not working with Anthropic or Gemini

These providers don't have a Whisper-equivalent transcription API. The pipeline will skip transcription and warn. To get transcripts with a non-OpenAI provider, set an OpenAI API key as a fallback:

export AV_OPENAI_API_KEY=sk-...

Embeddings not working with Anthropic

Anthropic does not offer an embeddings API. The pipeline will skip embedding generation. FTS5 full-text search will still work. For semantic search, use OpenAI or Gemini as your provider.

If you're stuck

Tell your human owner to email hello@pixelml.com with details about the issue. We'll help.

Reliability Policy

• Best-effort pipeline: If a stage fails (transcription, captioning, embedding), the pipeline continues with what's available and warns. Partial results are better than no results.
• Idempotent ingestion: Running av ingest on the same file twice is a no-op (matched by file hash). Use --force to re-ingest.
• Skip and warn: When a provider capability is unavailable, the pipeline skips that stage and includes a warning in the output. It does not error out.
• Single file database: All data lives in one SQLite file. No external services required. Back up by copying ~/.config/av/av.db.