Daemon

Your other self remembers.

A free, open-source memory for your Mac. Keep anything with one key, recall it by meaning with another. Everything runs on your Mac, with no account and no cloud.

Download for Mac · Build from source · How it works · Extend it

⌃⌥S keeps whatever is in front of you	⌃⌥Space recalls it by meaning, ↵ pastes it back
Hold ⌃⌥S to speak a voice note	Nothing selected? Drag over the part you want

Why Daemon

Useful things end up everywhere: screenshots on the Desktop, a line in a chat, a voice memo, a tab you meant to come back to. Daemon keeps them in one place and finds them again from a description like "the blue pricing table" or "the voice note about the caching idea", even when you never named the thing.

• One key, no decisions. ⌃⌥S does the obvious thing for what's in front of you (see the five rules).
• Search by meaning, across everything. Text, screenshots (and the words inside them), images, PDFs including scanned pages, Word and text files, audio, voice notes, video and links all live in one search space. A typed query finds a screenshot or a spoken sentence.
• Instant. Results on every keystroke in about 11 ms. ↵ pastes the result straight back where your cursor was.
• Private. Nothing leaves your Mac. One open model (EmbeddingGemma 2) runs locally on llama.cpp with Metal. No account, no analytics, no telemetry.
• Light. At rest: 0% CPU and ~21 MB for the app. The model process unloads its heavy parts when idle.
• Free and MIT-licensed. Fork it, ship it, build on it.

Install

Download

1. Get Daemon.dmg from the latest release and drag Daemon to Applications.
2. Open it. Release builds aren't notarized yet, so macOS will say it can't verify the developer the first time:
   • macOS 15 and later: System Settings → Privacy & Security → scroll down → Open Anyway.
   • macOS 14: right-click Daemon in Applications → Open → Open.
   • Or from Terminal: xattr -dr com.apple.quarantine /Applications/Daemon.app
3. On first launch Daemon downloads the model (~865 MB, once), shows a one-minute tour and asks for Accessibility.

Requires macOS 14 or later on Apple Silicon (M1 or newer).

Build from source

You need Xcode 26 or later.

git clone https://github.com/iamartyaa/daemon.git
cd daemon
make install    # release build → ~/Applications/Daemon.app, then opens it

Command	What it does
make build	Debug build of every target
make test	Unit, integration and UI tests (model-dependent tests skip until the model is downloaded)
make app	Release build/Daemon.app
make install	make app, install to ~/Applications, register with Spotlight and open
make dmg	build/Daemon.dmg; notarized when DAEMON_SIGN_IDENTITY and DAEMON_NOTARY_PROFILE are set

Tip

scripts/bundle.sh signs with your Apple Development certificate if one is in your keychain (a free Apple ID in Xcode is enough). Use one: macOS ties privacy permissions to the signature, so ad-hoc builds lose their Accessibility grant every time you rebuild.

Use

Shortcut	Action
⌃⌥S	Keep what's in front of you
Hold ⌃⌥S	Speak a voice note; let go to keep it
⌃⌥Space	Open Recall
↵	Paste the result into the app you were in
⌘C / ⌘↵ / Space	Copy · open in its app · Quick Look
⌘V in Recall	Keep whatever is on the clipboard
⇥ on a "Kept" toast	Add a note to what you just kept

Both shortcuts can be changed in Settings. You can also drop files on the menu bar icon.

The five keep rules

⌃⌥S checks these in order and does the first that applies:

1. Held down → a voice note
2. Text selected → the text, with the app and window it came from
3. Files selected in Finder → copies of the files (originals can move or disappear; files over 500 MB are kept by reference)
4. A browser with nothing selected → the page URL and title, plus a screenshot
5. Anything else → drag over a region to keep a screenshot

Recall syntax

Just describe it. Filters can go anywhere in the query and combine:

Filter	Examples
Source app	from:safari, from:figma
Kind	images, screenshots, voice, audio, video, pdf, links, type:text
Time	today, yesterday, this week, last month, in march, 3 days ago

from:arc last week pricing table → links and screenshots from Arc, last week, about pricing tables.

Privacy

• Nothing leaves your Mac. The only network requests are the one-time model download from Hugging Face and an optional update check (off in Settings).
• Daemon keeps only what you hand it. No background screen recording, no disk crawling, no clipboard history.
• Permissions are asked once, when first needed: Accessibility (read the selection, paste back) at onboarding; Screen Recording, Microphone and Automation (browser URL, Finder selection) the first time you use those features.
• Password managers are excluded by default (Passwords, Keychain Access, 1Password, Bitwarden, …). Add any app in Settings.
• Retention: kept items are deleted automatically after 7, 30 or 90 days (default 30), files and index included. Export everything to a folder at any time.

What	Where
Library (SQLite, vectors, keyword index)	~/Library/Application Support/Daemon/Library/library.sqlite
Kept files (deduplicated by SHA-256) and thumbnails	~/Library/Application Support/Daemon/Library/files/, thumbnails/
Model	~/Library/Application Support/Daemon/models/

To uninstall completely, quit Daemon, delete the app, then rm -r ~/Library/Application\ Support/Daemon.

How it works

flowchart LR
    K["⌃⌥S"] --> C["Capture<br/>5 keep rules"]
    C --> L[("Library<br/>SQLite + files")]
    L --> X["Extract<br/>chunks · OCR · PDF pages<br/>30 s audio windows · video frames"]
    X --> E["DaemonEmbedder<br/>EmbeddingGemma 2<br/>llama.cpp + Metal<br/>separate process"]
    E --> V[("768-d Float16<br/>vectors")]
    R["⌃⌥Space"] --> Q["Query parser<br/>filters + text"]
    Q --> E
    Q --> F["FTS5 keyword search"]
    V --> H["Hybrid ranker<br/>cosine ⊕ keyword, RRF<br/>relevance cutoff · recency"]
    F --> H
    H --> P["Recall panel → ↵ paste back"]

Loading

• One model, one space. EmbeddingGemma 2 maps text, images and audio into the same 768-dimension space, so text queries match pictures and speech directly. No captioning or transcription step to go wrong.
• The model lives in its own process. DaemonEmbedder talks JSON lines over stdin/stdout. A crash or memory spike never takes the UI down, and the process restarts after two idle minutes to release the image/audio encoders.
• Search is two-phase. Keyword hits come back within a few ms of each keystroke, meaning hits are merged in when the query embedding is ready, and stale answers are dropped. Vector search is brute-force cosine over Float16 with Accelerate (~12 ms for 100k segments): no ANN index to build, tune or corrupt.
• Capture never blocks. Selection and browser reads start on key-down off the main thread; toasts are Core Animation layers, so a keep costs 2–4% of a core.

Measurements, method and per-operation memory/CPU are in docs/qa-performance-report.md.

Project layout

Path	What's there
Sources/DaemonCore/Contracts.swift	The shared types and protocols every module talks through: Item, Segment, Capture, Embedder, CaptureSink, RecallDataSource. Start here.
Sources/DaemonCore/Library	GRDB/SQLite schema, content-addressed file store, export
Sources/DaemonCore/Extract	Item → segments: text chunking, PDF, OCR, audio/video windows
Sources/DaemonCore/Search	Query parser, Float16 vector index, hybrid ranker
Sources/DaemonCore/Embedding	EmbedderClient, which runs and supervises the model process
Sources/DaemonCore/App	Engine (the façade the app uses), indexing pipeline, Perf signposts
Sources/DaemonCore/Retention, ModelDownload, Stats	Auto-delete, resumable model download, local usage counts
Sources/DaemonEmbedder	The model process: llama.cpp + mtmd
Sources/DaemonCapture	Hotkeys, keep rules, selection/Finder/browser readers, region capture, voice, paste-back, permissions
Sources/DaemonUI	Recall panel, toasts, menu bar, onboarding, Settings, and Theme.swift (all colors, type, radii, timings)
Sources/Daemon	App entry point; AppController wires the modules together
Onboarding/remotion	The onboarding clips, written as Remotion code (npm ci && npm run render)
Sources/EmbedSpike, Spikes/MLXSpike	Embedder benchmark (make spike) and the MLX-vs-llama.cpp comparison
design/screenshots	Design reference: Ivory (light) and Smoked (dark)

Extend it

The modules only meet through the protocols in Contracts.swift, so most changes stay inside one folder.

Index a new file type

Extractor.segments(for:fileURL:workDir:) in Extract/Extractor.swift turns a kept item into SegmentDrafts, each with the text to keyword-index and an EmbedInput (.document, .image or .audio) to embed. Add a branch for your UTType and return drafts; the pipeline, storage and search pick them up. Add a fixture-based test in Tests/DaemonCoreTests/ExtractTests.swift.

Change what ⌃⌥S does in an app

ContextRules.decide(_:) in DaemonCapture/ContextRules.swift is a pure function from a ContextSnapshot to a KeepAction. Browsers that expose their URL over AppleScript are listed by bundle ID in the same file, as are the default excluded apps.

Swap the model

Anything that conforms to Embedder (embed(_ input: EmbedInput) async throws -> [Float]) can replace the default. For another GGUF model on llama.cpp, change the files in ModelDownload/ModelManifest.swift and the prompt prefixes in DaemonEmbedder/main.swift. Vectors record their modelVersion, so a new model means re-embedding the library.

Add a Recall filter

Search/QueryParser.swift turns words into SearchFilters; add a prefix or phrase there and its case in SearchQueryParserTests.

Restyle it

Every color, font, radius and animation timing comes from Theme.swift, with an Ivory and a Smoked value for each color. Views don't hardcode any of them.

Profile it

Hot paths are wrapped in os_signpost intervals (Perf.begin / Perf.end); open Instruments → Points of Interest. Benchmarks: DAEMON_PERF=1 swift test --filter RecallPerfTests and DAEMON_BENCH=1 swift test -c release -Xswiftc -enable-testing --filter "ToastPerfTests|PerfBenchmarks|OCRTests".

Contributing

Issues and pull requests are welcome.

1. Fork, branch from main (feat/…, fix/…), keep the change focused.
2. make test passes. UI changes: a before/after screenshot in both Ivory and Smoked.
3. Match the code around you. In UI copy the verbs are Keep / Kept and Recall, never "save" or "search".
4. Anything that touches capture or storage keeps the privacy promise: nothing leaves the Mac, nothing is kept that the user didn't hand over.

Good first areas: more browsers in ContextRules, more file types in Extractor, translations, and measurements on an 8 GB M1.

FAQ

Does it need the internet?

Only once, to download the model. After that it works fully offline.

How much disk and memory does it use?

The model is ~865 MB on disk. At rest the app uses ~21 MB and the model process ~160 MB. Keeping an image or audio loads the media encoders (~1–2 GB) until two idle minutes pass. Kept items take what the files take, deduplicated; each vector is 1.5 KB.

Intel Macs?

Not supported. The model runs on the Apple Silicon GPU through Metal.

Why doesn't it ask for Full Disk Access or record my screen?

It doesn't need to. Daemon only sees what you hand it with ⌃⌥S, a drop or ⌘V.

Acknowledgements

Daemon stands on EmbeddingGemma 2 by Google (Apache 2.0, downloaded at first run from ggml-org/embeddinggemma-2-GGUF), llama.cpp (MIT), GRDB.swift (MIT), KeyboardShortcuts (MIT) and Sparkle (MIT). The onboarding clips are made with Remotion, which has its own license for rendering.

License

MIT © 2026 Amartya Yadav