Recording and offline sync
The web client supports microphone capture in modern browsers. Chrome and Edge can also request display/system audio. That permission must begin from a user action and is requested again after a reload; unsupported or lost system audio is shown explicitly while microphone capture continues.
The macOS 13+ and Windows 10/11 x64 clients combine microphone and system audio into separate logical channels. Desktop is the v1 reference client for uninterrupted recording because browsers can suspend tabs and evict non-persistent storage.
The desktop mixer waits for both selected sources rather than consuming a temporarily late channel as silence. If a source actually stops, its aligned tail is finalized and the remaining source continues. Windows loopback audio is resampled from the active output-device format to 16 kHz mono before mixing.
Desktop quick capture
After sign-in, the desktop client opens as a compact always-on-top recorder in the lower-right corner. Microphone and device audio are selected together by default in both quick capture and the full recorder. Open notes library expands it into the same Record, Timeline, Memories, Search, Speakers, Sources, and Settings structure used by the web app. Closing either view hides NeoRecall to the tray so upload recovery and meeting detection can keep running; Quit in the tray menu exits it completely.
While the tray process is running, NeoRecall watches for newly launched meeting sessions from Zoom, Microsoft Teams, Webex, and FaceTime. A stable process transition surfaces the recorder; it never starts recording by itself. The user must acknowledge consent and press Record meeting, which captures microphone and device audio together. When the meeting process exits, an active recorder asks the user to finish instead of cutting audio off automatically.
This detector reads process metadata only—it does not monitor idle audio or record in the background. Browser-only meeting tabs cannot be identified from the operating-system process list, so Google Meet and other browser calls still use Quick capture from the tray.
Consent and indicators
NeoRecall shows a first-run recording notice and a persistent recording state. It has no covert mode. Before recording, obtain the consent required in your jurisdiction and organization. Pause immediately when consent changes or sensitive material should not be retained.
Recording context
Once capture starts, the idle source picker is replaced by one active-recording workspace. It contains the single live timer and stop control plus one action each for a highlight, typed note, photo, or file. Context is durably queued on the device alongside recording metadata, so adding it also works while offline.
Highlights emphasize the nearby transcript. Notes can supply names or details the audio may miss. NeoRecall extracts text from plain text, Markdown, CSV, JSON, PDF, and DOCX files, and asks a vision-capable configured model to describe JPEG, PNG, and WebP images. Other files remain attached and are reported as unsupported instead of being discarded. During consolidation, each item is associated with the closest transcript moment, so recordings that produce several memories do not attach every source to every memory.
The same Context section appears on a finished memory. Adding or removing a source there queues an in-place AI rewrite: the memory identity, pin/archive state, and explicit importance override remain intact while generated prose and mini-memories are replaced atomically.
Durable chunks
Capture uses independently decodable WAV chunks, normally 30 seconds long with a short leading overlap. The client keeps every unacknowledged chunk in its local ledger. When connectivity returns it:
- recreates the device, session, and source declarations idempotently;
- uploads chunks in source order with an idempotency key and SHA-256;
- polls terminal receipts if SSE is unavailable;
- deletes local audio only after the receipt includes transcript persistence, checksum, and server-audio-deletion timestamps.
NeoRecall never removes pending audio merely to satisfy a storage cap. If storage is exhausted, recording stops visibly and records a gap. Capture-gap records carry both monotonic time offsets and the exact missing sequence range. Later chunks remain blocked until every preceding sequence is terminal or explicitly covered by one of those gaps.
Local pending-audio totals, sessions, and uploads are account-scoped. Logging out first finalizes an active recording and never hands its queued chunks to a subsequently signed-in account.
Mobile capture foundation
Android phone-microphone capture uses a visible foreground service, an unbounded active wake lock, and a process-owned Flutter engine. Swipe-away does not destroy the recording engine. If Android later recreates the process while capture is still requested, NeoRecall closes the interrupted session and starts a new recovery session using the durable local intent. Network requests have bounded timeouts, and queued chunks remain on disk through offline periods.
Bluetooth support is transport- and device-adapter based. Adapters normalize PCM, connection states, battery state, and physical start/stop/standby/wake events. Reconnect uses bounded exponential backoff. Plaud Note Pro and NotePin S pair on iOS and Android as offline-first wearables: the pin records on its own, and NeoRecall drains finished files over BLE after a handshake token from this server. The consumer Plaud app cannot stay bound to the same device. Desktop and web clients can use other wearables over Bluetooth, but Plaud pairing stays on the phone. Memoket Gem pairs as a BLE wearable on the same clients: it streams live Opus and records on the device. Start and stop work from the app or the hardware button. Files recorded while disconnected are drained into the ordinary ingest pipeline.
Android cannot override a user Force stop, revoked permissions, exhausted storage, or some vendor battery-management policies. iOS cannot continue after a user force-quit. Those operating-system actions are shown as limitations, not described as guaranteed background execution.
Android home-screen widgets
Five widgets are available from the Android widget picker. All of them are configurable: long-press a placed widget and choose the gear, or use the settings screen shown while placing one. Every option previews live before it is saved.
| Widget | Shows | Configure |
|---|---|---|
| Recorder | Start capture, stop it again, and the elapsed recording time | What the button does while recording; appearance |
| Capture status | The current pipeline stage, its progress, queued audio, and any issue holding it up | Where a tap opens; appearance |
| Commitments | Open tasks and promises with their due dates | Which commitments to show; whether to show the source conversation; appearance |
| Memories | Recent memories with their summaries | Which memories to show; whether to show summaries; appearance |
| Today | One headline number against the last seven days, or the week ahead | Which number to headline; appearance |
Widgets render in the launcher's process, so they cannot reach the server or the local database. The app publishes one snapshot of what a widget may show, and the widget renders only that. A widget therefore shows the state as of the last time NeoRecall ran, which on Android is whenever capture, a wearable link, or sync is active.
Two actions do not open the app: stopping a recording, and ticking a commitment done. Both are recorded on disk before anything else happens, so a tap survives a cold start; a ticked commitment disappears from the widget immediately and is sent to the server the next time the app runs. Starting phone-microphone capture still opens NeoRecall, because Android refuses microphone access to a process with no attached UI.
Signing out clears every widget: they fall back to asking for a sign-in rather than continuing to show the previous account's memories.
Wear OS
NeoRecall for Wear OS is a separate APK that shares the phone app's application id and signing key. The watch records on its own microphone, in the same 30 second AAC chunks the rest of the pipeline uses, and holds every chunk in a local ledger until the phone returns a terminal receipt proving the transcript was persisted and the server audio deleted. A watch that is out of range keeps recording; nothing is deleted early, and the app screen says how many clips are being held rather than treating the backlog as a fault.
The watch has two screens. The first is the record control, the elapsed time, and what is still owed to the phone. Swiping through the day card opens the second: one scrolling page holding the day's totals, the newest conversation with its write-up and transcript, the memories that were kept, and the open commitments. The digest is pushed from the phone over the Wear Data Layer — the watch never contacts the server, and nothing it shows can change what the phone does with audio.
| Surface | Shows |
|---|---|
| App | Record control, elapsed time, held clips, phone link, and the full day digest |
| Record tile | Start and stop capture in one gesture from the watch face |
| Today tile | Talk time, memories kept, open commitments, and the newest headline |
| Recording complication | Whether NeoRecall is listening, and for how long |
| Today complication | What is overdue, due today, or open |
Neither complication takes an action: a watch face is read at a glance and pressed by accident, so starting or stopping the microphone is confined to the tile, which is looked at before it is touched.
The digest is trimmed to fit a Data Layer item: at most the last 40 transcript lines of the newest conversation, and a bounded number of memories and commitments. When lines are left behind the watch says how many are still on the phone rather than implying the conversation ends where the payload does.
Installing it
The watch app is not distributed through the Play Store. Each release publishes
NeoRecall-WearOS-<version>.apk, which is sideloaded onto the watch once:
- Download the APK onto a computer on the same network as the watch.
- On the watch, enable developer options, then ADB debugging and wireless debugging. The watch shows its IP address there.
adb connect WATCH_IP:5555, thenadb -s WATCH_IP:5555 install -r NeoRecall-WearOS.apk.- Open it once on the watch and allow the microphone.
Settings → Watch in the phone app lists every paired watch and says whether NeoRecall is installed on it, repeats these steps, and can push the current day to a watch that was just set up. Developer options can be turned back off afterwards.
Importing existing audio
Choose Import audio on the Record screen. Large files use the multipart import protocol and enter the same VAD, transcription, diarization, search, and memory pipeline as live capture. Selecting the same file again after an interrupted transfer resumes its missing parts for the same account and server. Failed imports expire according to the server TTL because a browser may no longer retain access to the original file.