DEV Community: Yisrael Gottlieb

The good open AI work is all in terminals. I put it behind a face you can talk to.

Yisrael Gottlieb — Wed, 23 Sep 2026 15:00:28 +0000

There is a lot of very good open work in this space right now, and almost none of it is something you can talk to. NVIDIA put out a speech model that runs on your laptop. An independent researcher put out an 82-million-parameter voice. Nous Research put out an agent runtime under MIT. Every one of those is excellent, and every one of them ends at a terminal or a Python import.

Memoe is those pieces, assembled, behind a face you can talk to. No chat box: it lives in your notch, in every app on your Mac. You say "Hey Moe", it answers out loud, and it does the work. https://dub.sh/nxzivCr

Here is what is in it and where each part comes from.

Speech, both directions, on the Neural Engine

Speech to text is NVIDIA's Parakeet (parakeet-tdt-0.6b-v3), converted to CoreML and running on the Apple Neural Engine. End-of-utterance detection is a second, smaller Parakeet streaming model, so Moe knows you have finished a sentence without waiting on a timer. Speaker diarization is in there too, which is what makes meeting notes name the right person.

The voice that answers is Kokoro, 82 million parameters, also CoreML on the ANE.

Nothing here is a cloud call. That is why it is free, why it is fast enough to interrupt, and also why Memoe is Mac-only. This does not work as an Electron app.

The harness is Hermes

The agent runtime is Hermes Agent from Nous Research, MIT licensed: model routing, tools, memory, the plugin surface. It runs as its own process that Moe launches and talks to over HTTP on 127.0.0.1, in a URLSession with the proxy dictionary explicitly emptied, so a system-wide proxy never sees your mail or your memory.

Hermes is genuinely good and almost nobody has met it, because meeting it means a terminal.

The model is the one you already pay for

Memoe brings no model of its own. It drives the official Claude Code or Codex CLI as a long-lived child process, or an API key, or a local model in LM Studio or Ollama.

So if you already pay for Claude or ChatGPT, Memoe is free and you get exactly the model you already trust. Same quality, out of the chat window. One subscription, not two.

What I actually built

The part where those stop feeling like pieces.

A window that is not a window: the island under the notch, which grows out of the notch when Moe is thinking and shrinks back when it is done. Wake word and push-to-talk that work from any app without stealing your cursor, so you can keep typing somewhere else while it works. Turn-taking that does not cut you off at a breath. A memory that is one plain text file on your disk, where every line says where it came from, so you can open it, fix a line, or delete a line. A ladder for driving the Mac that goes structured data, scripting, menus, keyboard shortcuts, accessibility elements, and only then pixels. Connections to mail, calendar and messages that are already set up when you first speak to it.

And two rules that came before any feature: before Moe sends anything as you it shows you what is going out and waits for a yes, and if you touch the trackpad while it is working it stops.

Try it

macOS 15 or later, Apple silicon. Signed, notarized, updates itself. Free.

https://dub.sh/nxzivCr or brew install --cask izzy-gottz/moe/moe

It is beta. If it breaks, tell Moe and it reaches me.

Giving an AI assistant a memory you can actually read

Yisrael Gottlieb — Mon, 14 Sep 2026 15:00:59 +0000

Moe is an assistant that runs on your Mac all day. You say "Hey Moe" from whatever app you're in, ask for something in a normal sentence, and it does it. The part of it I've thought about most is memory, so this post is about how that works.

A short list, in plain files

What Moe knows about you lives in Markdown files on your disk, under ~/Library/Application Support/Moe/memory. It starts with four notes you can edit yourself (About me, People, Projects, Preferences), and Moe adds to them as you talk.
The list is small on purpose. Every word of it is read before Moe answers anything, so a line has to earn its place.

Every line has a source

A line in Moe's memory looks like this:
Marcus - the printer; wants the print files with every export Messages · 14 Jun · highThat's the fact, where it came from, when, and how sure Moe is. Things it picked up in conversation are marked as coming from your conversations. Things it learned by reading your mail, messages, calendar or browser carry the app and the date, plus a confidence: sure, fairly sure, or not sure.
The source is the part I care about. When an assistant gets something wrong about you, the first thing you want to know is why it thinks that, and the second is how to fix it.

Fixing it

The Memory page in the app shows every line, grouped into things about you and things about the people and projects around you. Each one has a Fix or forget button. Correct a line that was learned from reading and the change sticks, so the next pass over the same mail won't quietly bring the old version back.
Lines from Moe's own conversation memory work differently. Moe writes that file continuously while you talk, so if the app edited it behind Moe's back the two would overwrite each other and your fix would lose. The button sends your correction to Moe in chat instead, and you watch it make the change.
Sometimes Moe notices something it can't place. Those show up as questions at the bottom of the page, and you answer them in your own words.

Where it learns from

Moe learns from your conversations as they happen, and once a day from whatever you've let it read. All of it stays on your Mac.
Moe is free on the Mac and uses the AI you already have: Claude, ChatGPT, an API key, or a local model. It's in beta. moebot.app

I built an AI assistant that lives on my Mac instead of in a chat tab

Yisrael Gottlieb — Sat, 12 Sep 2026 15:02:03 +0000

Every AI tool I used worked the same way. I'd open a tab, explain who the client was and what we'd agreed, paste in the email, get some text back, and then go do the actual work myself. The model was smart. The setup around it left all the carrying to me.
So I built Moe, an assistant that runs on your Mac the whole time you're using it. This post is about the two decisions that shaped it: how you reach it, and how it does things once you've asked.

Reaching it from any app

Moe has no window you have to switch to. It sits in a small panel at the notch, over whatever you're doing. There are three ways to start talking:

Hold the right Option key, talk, let go.- Press Control-Option-Space to start and stop hands free.- Say "Hey Moe". The wake word runs on a small model on your Mac and listens only for its name.After it answers it keeps listening for a few seconds, so a follow-up doesn't need the key or the name again. Talk over it and it stops. If you change one part of what you asked halfway through, it keeps the rest. The hard part was working out which speech was meant for it. Moe looks at how long ago it spoke, whether it just asked you something, how loud you are compared with the room, and whether the voice matches the person who started the conversation. Someone talking beside you doesn't get answered. ## Doing the thing Once you've asked, Moe does the work on the Mac. It uses the same accessibility APIs a screen reader uses, and clicks go straight to the target window in the background. Your cursor stays where it is and you can keep typing in another app. Clicking is the last thing it tries, though. Asked to join a meeting, it reads the link off the calendar event and opens it. The order is: data it already has, then the app's own scripting, then a menu item or keyboard shortcut, and only then a click. Reading an app's menu bar directly turned out to matter most. It's faster than looking at the screen, and it works in apps that draw everything on a custom canvas, like audio and 3D tools, because their menus are still ordinary menus. It doesn't ask before every click. It asks when something is about to leave or can't be undone: paying, sending a message or an email, deleting, typing a password. It shows you what's going out and to whom, then waits. Touch the trackpad while it's working and it stops. ## It remembers you Moe has read the mail, messages and calendar you connected, and it keeps a short list of what it has learned about you in plain Markdown on your disk. Every line says where it came from. I'll write up how that works in another post. ## What it runs on I didn't build the hard parts. Speech recognition is NVIDIA's Parakeet running on the Mac, and the agent runtime is Nous Research's Hermes Agent. The thinking comes from whatever you already use: a Claude or ChatGPT account, an API key, or a model on your own machine. Moe doesn't resell it or mark it up. It's free on your Mac and needs macOS 15 on Apple silicon. It's in beta, so you'll hit rough edges. If something breaks, tell Moe "that didn't work" and it reaches me. moebot.app