TL;DR

• Private, on-device search across your screenshots, memes, and photos.
• OCR + fuzzy text search today; hybrid semantic search next.
• Nothing ever leaves your device. Fast after initial indexing.

Links

• Demo (30s)
• GitHub: https://github.com/zmaas/memeory

Why

Finding “that one screenshot” is annoying. Memeory lets you type what you remember and instantly jump to the right image, completely offline.

Highlights

• Fully local pipeline: scan → OCR → index → search. No servers, no tracking.
• Fuzzy text search over OCR output (handles typos and partial matches).
• Share or open the matched image directly from results.

How it works (v1)

1. Permissioned gallery scan with content URI access.
2. OCR via ML Kit → automatic caption generation
3. Fuzzy search (Fuse.js) over all captured text.

Roadmap (v1.1+)

• Hybrid search: text + semantic embeddings with an ANN index.
• Model: small image encoder (quantized SigLIP or similar)
• Prototyping integration with ExecuTorch/Expo

Privacy

• Fully offline. No external calls; no analytics by default.

Tech

• React Native (Expo), TypeScript
• ML Kit Text Recognition
• ExecuTorch (for embeddings, v1.1, in development)

What I learned

• Android media pipeline quirks, content URIs, and batching I/O for speed.
• OCR normalization matters more than model size for real-world recall.
• Fuzzy search + OCR gets you 80% of the way; semantic adds the “vibes.”

Demo