Show HN: AI search for every photo and every frame of video on macOS

SCM (Screen Memories) is a local-first macOS app that indexes every photo and video frame on your machine so you can search by plain-language memory descriptions. It builds embeddings for whole files and per-scene segments, extracts OCR from visible text, and transcribes spoken words with Whisper so searches can land on an exact shot or spoken line. Searches run offline after an initial model download, support five modes (Files, Scenes, OCR, Dialogue, optional local LLM chat), surface why each result matched, and let you save queries as persistent tabs. Library maintenance is automatic: watched folders import, file-hash deduplication handles renames, and model switches re-embed in the background without blocking lookup. Special tabs detect screenshots and images containing email addresses with tolerant OCR reassembly.

Under the hood it uses ONNX vision models (default CLIP ViT-L/14@336 plus several SIGLIP variants) with cosine-similarity ranking, gated filename/phrase boosts, an honesty floor, and near-duplicate diversity filtering. Video shot detection is done with ffmpeg and configurable segment-density presets; OCR uses Tesseract (English + ~35 languages); dialogue retrieval is literal over Whisper transcripts with tiered exactness; optional local chat runs a llama.cpp sidecar (QWEN3 1.7B default). Built with Electron and Bun, it ships as a macOS app (Homebrew tap available) and provides developer build/package commands for signing and distribution.