No chat box. Talk from any app, and it talks back.

Memoe (said MEM-oh, like memo) is a voice assistant that lives in your Mac's notch.
Already pay for Claude or ChatGPT? Then Memoe is free.

Download · Docs · Build from source · Discussions · Contributing

https://github.com/user-attachments/assets/3f1d9316-e605-4fd4-9b55-ad303be67e62

_{Sound on. A rendered film of real features; Moe's voice is the app's own.}

Download

Download Memoe for Mac: a Mac with Apple silicon and macOS 15 or later.
Or with Homebrew:

brew install --cask izzy-gottz/moe/memoe

It is a large download, a little over a gigabyte, because the speech and voice
models and the part of Moe that thinks are all inside it. Drag Memoe onto
Applications and open it. Memoe is signed and notarized, so macOS asks once
whether to open an app from the internet: click Open.

Moe thinks with an account you already have: a Claude or ChatGPT plan, an API
key from any model provider, or a model running on your own Mac (Ollama, LM
Studio or another local server). There is no Memoe model and no bill for
thinking. See Bring your own brain.

Prefer to build it yourself? Jump to Build from source.

What it does

Lives in the notch, not a window. Hold right ⌥ in whatever app you
are in, or double-tap it, and talk. ("Hey Moe" works too, once you turn it
on in Settings.) Moe answers out loud and on a small panel under the notch,
and you can interrupt it at any time. A Mac without a notch gets a small
black tab in the menu bar instead.
Does the work on your Mac. It opens apps, uses their menus, fills in
fields and uses its own browser for the web. It tries the app's data and
scripting first and clicks only as a last resort, in the background, so your
cursor does not move.
Remembers you. A short list of plain facts about you that Moe reads
before every answer. You can read every line, correct it or delete it, on the
Memory page.
Asks before anything goes out. Before an email, a message, an
invitation or a shared file goes to another person, a sheet shows where it
is going, to whom, and the whole text, which you can edit. Nothing is sent
until you press Send.
Takes the meeting notes. It records both sides of a call, transcribes
it and separates the speakers on your Mac, then writes the notes. No bot
joins the call, and recording never starts without you.
Reaches your services. Mail, Calendar, Contacts, Notes and Reminders
on this Mac, plus Gmail, Slack, Notion, GitHub and many more through
Composio or any MCP server you add.

Every screen and every switch is documented at docs.memoe.app,
built from doc.md in this repository.

Bring your own brain

Moe has no model of its own. It thinks with an account you already have:

a Claude or ChatGPT plan, signed in through your browser;
an API key from any model provider;
or a model running on your own Mac (Ollama or another local server).

Pick one in Settings › Brain and change it whenever you like. We put no
model of our own behind it and never bill you for thinking.

How it works

flowchart LR
  subgraph Mac["Your Mac"]
    App["Memoe.app (Swift)<br/>notch, voice in and out, memory,<br/>send sheet, meetings"]
    Hermes["Hermes Agent<br/>(separate process)<br/>agent loop, tools, skills, schedules"]
    Cua["cua-driver<br/>Mac control through<br/>the accessibility tree"]
    Local["On-device models<br/>Parakeet, Kokoro, Whisper,<br/>wake word, turn detection"]
    App -- "local HTTP" --> Hermes
    App --- Local
    Hermes -- "MCP" --> Cua
  end
  Hermes -- "your plan or key" --> Provider["Your AI provider<br/>(or a model on your Mac)"]
  Hermes -- "MCP and APIs" --> Services["Services you connect"]

Memoe.app is a native Swift app with no Dock icon. It listens, speaks,
shows the panel under the notch, keeps your memory, and gates anything that
goes out.
Hermes Agent by Nous
Research (MIT) runs the thinking loop. Memoe installs a pinned commit of
a fork with Memoe's changes;
the pin is in scripts/install-hermes.sh.
cua-driver (MIT) drives the Mac through
the accessibility tree.
Speech runs on FluidAudio
and WhisperKit on Core ML.

Privacy: what leaves your Mac

Speech recognition, the voice, meeting transcription and Moe's memory run on
your Mac. Your data lives in ~/Library/Application Support/Moe.

The official download. In the words of the docs:

On the free plan very little reaches us, and this is all of it. What you say
to Moe, and what Moe reads to answer you, goes to the AI provider you chose
to think with (your Claude or ChatGPT plan, or the company whose key you gave
it), or stays on this Mac if you use a model running here. Services you
connect through Composio go through Composio, and on the way they pass
through our server, which holds the Composio key, counts each request
against your plan's limits, and does not store what the requests say. Three
things send something to us, and each one asks you first: signing in with
Google, which sends only who you are; starting Away, which moves Moe into
your own cloud computer; and problem reports, which send what you asked us
to look at. Two more happen on their own, and there is no switch for them:
once a day Moe asks our website whether there is a newer version, a plain
request that says which version of Memoe is asking; and at most twice, ever,
it sends our server a single word to count a person coming back, once if a
third spoken command gets a real answer in your first week and once the
first time one does on or after day 7. That word is all it carries: no
account, no device or Mac identifier, and never what was said. A Memoe you
build yourself from the open-source code sends neither count and does not
check for updates. The speech recognition, the voice, the transcription and
the memory all run on your Mac. The full policy is at
memoe.app/privacy, and it lists every service
that ever touches your data.

A build from this repository has no code for our cloud services:

What	Where it goes	When
What you say to Moe, and what it reads to answer you	The AI provider you chose (your plan or key), or nowhere if you use a model on your Mac	Every request: this is the thinking, and it goes to the account you picked, not to us
Sign-ins to services you connect through Composio	Composio, under your own Composio account	Only for services you connect that way
What you ask Moe to send	The person or service you approved on the send sheet	Only after you press Send

What is open, and what is not

	This repository	The official download
The Mac app and everything it runs on your Mac	✓	✓
Sign-in, the Composio proxy, problem reports	—	✓ (closed)
Away, the optional paid cloud service	—	✓ (closed)
Updates itself	—	✓
Signed and notarized by the maintainer	—	✓
Price	free	free; Away is a paid plan

The closed parts are not in this repository, and a build from this repository
does not need them. The official download is built from this code plus those
closed parts.

Build from source

You need a Mac with Apple silicon, macOS 15 or later, and the Xcode Command
Line Tools (xcode-select --install). There is no Xcode project.

git clone https://github.com/memoe-app/memoe
cd memoe
./scripts/bootstrap.sh            # builds the community edition and opens it
./scripts/bootstrap.sh --models   # the same, with the speech and voice models bundled (~900 MB)

A build from this repository is the community edition, named Memoe OSS:
its own bundle identifier (app.memoe.oss), its own data folder
(~/Library/Application Support/Memoe OSS), no updater and no cloud. It can
never take the place of an official Memoe installed on the same Mac. Without
--models, each model downloads the first time it is needed. On first launch
the app installs Hermes from the pinned commit of the fork.

Tests: MOE_SKIP_PG=1 ./scripts/test.sh runs the Swift tests.

Build inputs that are not in git

Path	What	How you get it	Required?
`Models/`	Core ML speech and voice models	`./scripts/bootstrap.sh --models`, or `scripts/fetch-models.sh` (from Hugging Face)	No: without them each model downloads when first needed
`hermes-seed/`	Hermes, prebuilt so the app thinks offline on first run	`scripts/stage-hermes-seed.sh`	No: without it the app installs Hermes on first run
`Resources/google-client.json`	A Google OAuth client of your own	Google Cloud Console	No
`Resources/telegram-client.json`	Telegram app credentials of your own	`scripts/stage-telegram-client.sh`	No

Composio in a build from source

Composio is optional. Without it, everything on your Mac and any MCP server
you add still works.

The official app reaches Composio through Memoe's own server, which holds the
Composio key, so the Mac never does. A build from this repository has no such
server. It talks to Composio directly, with a key from your own Composio
account (platform.composio.dev). Put it in the
Keychain item the app reads, then quit and reopen the app:

security add-generic-password -s moe-composio-scoped -a api-key -w   # paste your ak_… key when asked

The app copies it into Hermes's .env (where Hermes's MCP client reads it) and
starts a Composio session with it. Connections and their sign-ins then live in
your Composio project, and Composio's usage limits and pricing are between you
and Composio.

A build from source is signed with your own identity (or ad hoc), uses a
different identity from the official app, and does not update itself.

FAQ

Do I have to pay for anything?
No. The app is free, and Moe thinks with an account you already have. The only paid thing is Away, an optional cloud plan, which is not in this repository. Can it run entirely on my Mac?
Speech recognition, the voice, meeting transcription and memory always do. Point Settings › Brain at a model on your own Mac (Ollama, LM Studio or another local server) and the thinking does too. My Mac has no notch.
Moe gets a small black tab in the middle of the menu bar instead. Intel Mac?
Not supported: Memoe needs Apple silicon and macOS 15 or later.

Contributing

Bug reports, fixes and ideas are welcome. Please read
CONTRIBUTING.md first. In short: open an issue for a bug you
can reproduce, discuss larger changes before writing them (in
Discussions or an issue), keep
pull requests small, and sign off your commits (git commit -s): this project
uses the Developer Certificate of Origin.

Found a security problem? Please do not open an issue; see SECURITY.md.

Credits

Memoe did not build the hard parts, and says so.

Hermes Agent by Nous Research: the thinking loop (MIT)
cua-driver by Cua: Mac control (MIT)
FluidAudio by FluidInference: speech recognition and synthesis (Apache-2.0)
WhisperKit by Argmax: Whisper on Core ML (MIT)
Sparkle: updates (MIT)
ONNX Runtime: wake word and turn detection (MIT)
OpenWhispr: transcript cleanup prompts, adapted (MIT)
Models: NVIDIA Parakeet TDT 0.6B v3 and TDT-CTC 110M (CC-BY-4.0), Parakeet Realtime EOU 120M (NVIDIA Open Model License), OpenAI Whisper large-v3-turbo (MIT), hexgrad Kokoro-82M (Apache-2.0)

The complete list, with every licence and attribution, is in NOTICE
and in the app under Settings › About & Licences.

License and trademark

Memoe is licensed under the Apache License 2.0. Third-party
components keep their own licences; see NOTICE and
LICENSES.md. Model weights are under their own licences and are
not part of this repository.

"Memoe", "Moe" as the name of this assistant, and the Memoe logo and icon are
trademarks of Yisrael Gottlieb. The licence covers the code, not the names or
the logo. If you distribute a modified build, give it a different name, icon
and bundle identifier. See TRADEMARK.md.

The service icons in Resources/ServiceIcons/ are trademarks of their owners,
used only to label those services in the app, and are not covered by this
repository's licence.

memoe.app · docs · [email protected]