A new open-source project is betting that the fastest way to find a memory is to describe it — and that the search should never leave your laptop. SCM, short for Screen Memories, is a free macOS application that applies local AI models to every photo and every frame of video in a folder, letting users search their libraries in plain language. The project appeared on GitHub this week and quickly climbed to the front page of Hacker News, where it drew more than 85 points within hours.
The pitch is simple on the surface: point SCM at any folder, and it builds a searchable index of what is actually in your media — not just filenames, but the visual content of each image and each scene inside each video. What sets it apart in a crowded field of AI photo tools is that all of it runs on the Mac itself. There are no accounts, no cloud uploads and no telemetry. Once the models download their weights on first run, the app works offline. For more context on this story, see our ongoing AI trends.
Five Ways to Search a Lifetime of Media
SCM organizes search into five distinct modes, each powered by a different model running locally.
The Files mode performs whole-file semantic search across photos and videos, ranking results by meaning with boosts from filenames and phrases. Scenes goes deeper: the app segments videos into individual scenes and embeds each one separately, so a query lands you on the exact shot, with a jump straight to its timecode rather than a timestamp buried in a filename.
Two text-focused modes handle the words inside media. The OCR mode, built on the Tesseract engine, reads text that is visibly present in images and video frames and matches it literally, with English plus 35 additional language packs that can be toggled on. The Dialogue mode runs Whisper speech recognition over videos and indexes the exact spoken words, with tiered levels of match exactness for hunting a specific line.
Finally, an opt-in LLMs mode turns the extracted material — dialogue transcripts, OCR text and filenames — into a local chat context. Users can ask questions about their own library, and the app returns answers with citations pointing back to the source files, all generated by a model running on the same machine.
How It Works Under the Hood
Technically, SCM is an Electron application written in JavaScript, using the Bun toolkit as its package manager and runtime. The heavy lifting is done by transformer models executed directly in the app — the default vision model is a CLIP variant whose weights, roughly 435 megabytes, are downloaded once on first use.
The index maintains itself. Watched folders auto-import new media as it arrives, content hashing deduplicates files that have merely been renamed or moved, and switching to a different model re-embeds the library in the background without blocking searches. The developer notes that menu-bar and tray features are specific to macOS, where the app is packaged for distribution.
The privacy model is the selling point. The project's documentation is explicit: media never leaves the machine, inference runs locally through transformer runtimes, and after the initial weight download everything works offline. For users with sensitive archives — medical imagery, legal footage, personal video — that architectural choice is the feature.
Built for Personal Archives
The interface reflects the same philosophy of customization. Any query can be saved as a reusable tab, so a search that took a moment to phrase becomes a permanent view over the library. The app also ships with toggleable Screenshots and Email tabs, acknowledging that a modern Mac accumulates screenshots and attachments at a pace that quickly outgrows the operating system's own search.
The underlying technique is the now-standard machinery of semantic search: a vision model converts each image and video scene into a numerical embedding, and queries are embedded the same way so that matching happens by meaning rather than keyword overlap. What is notable here is not the technique but the packaging — the same pipeline that cloud photo services run on server farms is compressed into an Electron app that a non-developer can install and run unattended over folders of their choosing.
Part of a Larger Shift Toward Local AI
SCM arrives as running capable AI models on consumer hardware has gone from novelty to expectation. Apple has steadily pushed semantic search into its own Photos app, while open-source speech, vision and text models have shrunk to sizes that run comfortably on laptops. What remains scarce is not the underlying models but the integration layer — software that wires them together into something a non-engineer can actually use on their own files.
That is the niche SCM occupies, and its reception suggests the itch is real. The repository collected 160 stars and eight forks in roughly a day, with Hacker News commenters focusing their questions on model sizes, indexing speed and Windows availability — currently the answer to the last one is no, as the project is macOS-only for now.
The project is released under the MIT license, which permits commercial use and modification. For anyone with a large, disorganized archive of photos and screen recordings, and a principled dislike of uploading it anywhere, it is one of the more compelling demonstrations yet that a fully local media search engine is no longer a research demo — it is an afternoon install.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
