What are Memory Notch's system requirements?

macOS 26 or later on an Apple Silicon Mac. macOS 26 is required because the live transcription path uses SpeechAnalyzer and SpeechTranscriber, which do not exist in earlier versions. A notch is not required: Macs without one get a capsule tucked under the menu bar.

How much disk space do the models need?

About 1.7 GB once every feature has been used: 920 MB for live transcription, 634 MB for the enhanced re-transcription, 98 MB for proper nouns, 42 MB and 13 MB for speaker labelling, and 1 MB for voice activity detection. Recordings are stored separately, in ~/memory.

Can I change which models Memory Notch uses?

Not today. Each feature is pinned to a specific model variant, and Settings lets you download those models ahead of time rather than swap them. What you can change is the vocabulary file of names and jargon, the input microphone, and whether a recording also captures the screen.

SPECIFICATIONS

What it runs, and what it needs

Sizes below were measured on disk rather than estimated. Everything here describes the app as it stands; where a number is not something we have benchmarked, it says so instead of guessing.

Requirements

Operating system	macOS 26 or later. The live transcription path is Apple's `SpeechAnalyzer` and `SpeechTranscriber`, which do not exist before 26. There is no build for an earlier macOS.
Processor	Apple Silicon. The app ships as an arm64 binary, and transcription and speaker separation run on the Neural Engine. Intel Macs are not supported.
A notch	Optional. On a Mac without one the panel is a capsule tucked under the menu bar instead, and everything else is the same.
Network	Needed once, to fetch the models. After that a recording transcribes with the network off. A signed release also checks a static URL for a version number daily, which Settings can switch off.

The models

Four features, each with its own model, all running locally. Sizes are what they occupy on disk after downloading.

Feature	Model	On disk
Live transcription	Nemotron multilingual streaming, 1120 ms chunks	920 MB
Enhanced transcript	Nemotron multilingual, 2240 ms chunks	634 MB
Proper nouns	Parakeet CTC 110M	98 MB
Speaker labels, streaming	LS-EEND	42 MB
Speaker labels	Diarizer: segmentation and embedding	13 MB
Voice activity	Silero VAD	1 MB
Total	Once every feature has been used	about 1.7 GB

They download the first time each feature is used. Settings lists them individually with a Download button, so you can fetch them on wifi rather than wait mid-recording.

Can the models be changed?

Not today. Each feature is pinned to a specific model variant, and the Models pane in Settings downloads them rather than swapping them. If you want a different speech model, this is not yet the app for that.

What you can change: the vocabulary file of names and jargon, which input microphone is used, and whether a recording also captures the screen.

Memory

Measured, not estimated: peak resident memory of one process, three runs each, transcribing a sixteen-second two-voice recording on an Apple Silicon Mac.

Doing what	Peak memory
Sitting in the notch, nothing running	16 MB
Live transcription	33 MB
Working out who spoke	117 to 121 MB
Transcribing and labelling together	121 MB
The same with the multilingual model	188 to 191 MB once warm, about 600 MB on the first load after boot

The number worth noticing is that the models take about 1.7 GB on disk and a few hundred megabytes of memory. CoreML memory-maps the weights and the Neural Engine reads them from there, so they never all land on the heap at once. Live transcription is the lightest of the lot because those models are managed by macOS and run outside the app entirely.

Disk is the real cost of this app. Memory is not, and any Apple Silicon Mac has enough of it.

Permissions it asks for

Microphone, the first time you record.
Speech recognition, for the on-device speech models.
Screen Recording, only if a recording needs system audio or the screen. Until it is granted, recordings are microphone only, and a fresh grant needs the app relaunched before it takes effect.

What lands on disk

One directory per recording, under ~/memory, holding everything that recording owns.

~/memory/notes/2026-09-10-0941-note-id/
  2026-09-10-0941-note-id.md     note and frontmatter
  2026-09-10-0941-note-id.txt    transcript
  2026-09-10-0941-note-id.m4a    audio
  2026-09-10-0941-note-id.mov    screen recording, if there was one

Alongside them sit README.md, which describes the layout for anything reading the directory, and vocabulary.md, which is yours to edit. There is a command line too: memory status, memory ingest recordings, memory migrate and memory delete <id>.

What it is good at

One-to-one calls, where system audio puts the other person in the transcript.
Conversations in a room you are part of, rather than meetings you send a bot to.
Interviews and consultations where the recording is the sensitive thing.
Anything you would hesitate to upload, which is the case the whole design is built around.

What it is not for

Team meeting archives. There is no shared workspace, no CRM sync and no cross-meeting analytics. Fireflies and Otter are built for that.
Clinical documentation. No BAA, no EHR integration, no HIPAA claim. See the ambient scribe page.
Anything but a Mac. No Windows, Linux, iOS or web version.

Known limits

Two people on one laptop microphone is the hardest case for any speaker separation, and sometimes comes back as a single speaker.
Speakers rather than headphones. The microphone hears what the speakers play, so a phrase can occasionally appear twice. Most duplicates are removed automatically; headphones avoid it entirely.
The display has to stay awake for system audio to keep arriving. The microphone keeps recording either way, and the note records where the far end cut out.
On AirPods, choose the built-in microphone. Used as a microphone, AirPods switch to a low-rate call mode that stops system audio being captured.

How it compares · See the demo