jayzes-memorex

(Yes, I know this is technically a Maxell ad, not a Memorex one. But it's a better image.)

Memorex

Is it live, or is it Memorex?

Give Claude perfect recall of any video. Memorex converts video and audio into structured markdown with transcripts and keyframes—so Claude can see what you saw and hear what you heard.

Why Memorex?

Claude can't watch videos. But it can read markdown and view images. Memorex bridges that gap:

• Transcripts with timestamps — Every word, synced to the timeline
• Smart keyframes — Only the frames that matter, not 30 fps of redundancy
• One command — Point it at a video, get back something Claude understands

Demo

$ memorex demo.mp4

memorex
  Processing: demo.mp4
  Duration: 2m 34s

✓ Extracted 154 frames
✓ Found 12 keyframes
✓ Keyframes saved
✓ Audio extracted
✓ Transcribed 47 segments
✓ Markdown generated

✓ Output: demo_memorex.md
  Frames: demo_memorex_frames/
  Estimated tokens: ~15,600

Install

Prerequisites: FFmpeg and whisper.cpp

# macOS
brew install ffmpeg whisper-cpp

# Ubuntu/Debian
sudo apt install ffmpeg
make install-whisper  # builds whisper.cpp from source

Install memorex:

go install github.com/jayzes/memorex/cmd/memorex@latest

The Whisper model (~148MB) downloads automatically on first run.

Usage

memorex video.mp4                    # Basic usage
memorex -t 0.9 presentation.mov      # Fewer keyframes (fast-changing video)
memorex -t 0.7 interview.mp4         # More keyframes (static video)
memorex --no-frames podcast.mp3      # Audio only
memorex --no-transcript silent.mp4   # Video only
memorex -q 20 -s 0.3 huge.mp4        # Smaller output

Options:

Flag	Default	Description
-o, --output	<input>_memorex.md	Output path
-t, --threshold	0.85	Frame similarity (lower = more keyframes)
-q, --quality	30	JPEG quality (1-100)
-s, --scale	0.5	Frame scale factor
--no-transcript		Skip transcription
--no-frames		Skip frame extraction

Output

video_memorex.md
video_memorex_frames/
├── frame_0001.jpg
├── frame_0015.jpg
└── frame_0089.jpg

The markdown gives Claude everything it needs:

# Video Analysis: video.mp4

## Metadata
- Duration: 2m 34s
- Keyframes: 12
- Token estimate: ~15,600

## Transcript

[0:00] Welcome to this demonstration...
[0:15] As you can see on screen...

## Keyframes

### Frame 1 (0:00)
![Frame at 0:00](video_memorex_frames/frame_0001.jpg)

Claude Code Plugin

Let Claude handle everything automatically.

/plugin marketplace add jayzes/memorex
/plugin install memorex@jayzes-memorex

Then just ask Claude to analyze a video. It'll run memorex, read the output, and tell you what it sees.

How It Works

1. Extract — FFmpeg pulls frames at 1 fps
2. Compare — Normalized cross-correlation finds visually distinct frames
3. Transcribe — whisper.cpp converts speech to timestamped text
4. Package — Everything becomes Claude-readable markdown

Development

make test    # Run tests
make lint    # Run linter
make fmt     # Format code

Troubleshooting

Problem	Solution
FFmpeg not found	brew install ffmpeg or apt install ffmpeg
whisper-cli not found	brew install whisper-cpp or make install-whisper
Out of memory	Use -s 0.25 -t 0.95 for large videos

Contributing

Fork, branch, code, test, PR. The usual.

License

MIT