Questions about Meringo Minutes

Yes, once the models are downloaded. By default, recording, transcription and the summary all run on your Mac. The internet is needed only to download each model the first time.

No, and I'd rather say so plainly. Recordings and meeting files are ordinary files in a folder in your Library, and the app adds no encryption of its own. macOS file permissions protect them, and FileVault does too if it's on. By default nothing is uploaded, so there's no server copy to protect.

By default, no. The only traffic is models coming in the first time they're needed, and there's no analytics, telemetry or crash-reporting code in the app. If you point the summariser at Ollama on another machine, each transcript goes there, and the privacy panel turns orange and names it.

Not in the current build, which never contacts an update server. When I add an update check, the privacy page will say what it sends, in the same release.

In ~/Library/Application Support/Meringo Minutes/, one folder for everything. Deleting a meeting moves it to the Trash, so you can get it back until you empty it.

No. It hears the other side of a call through a Core Audio tap, so it never asks for Screen Recording.

Yes: point the server URL in Settings at another machine running Ollama. Each transcript is then sent there, and the privacy panel turns orange and names the machine. The audio and the transcript stay stored on your Mac.

Your microphone, the audio your Mac plays during a call, and your calendar, to name meetings. Reminders when you send a commitment there, and your own Contacts card only when you ask. The signature holds exactly three entitlements: audio input, address book and calendars.

No cookies and no analytics. It keeps your theme choice in your browser. GitHub, which hosts the site, logs IP addresses for security. The privacy page has the details.

It gets about one word in five wrong: an 18.4% median word error rate across 66 bench recordings. In the demo it heard “locum” as “locker” and “agreed” as “gag read”.

As far as you can check it. Most summary sentences carry a timestamp, and with Ollama a second model sets aside what it can't support. Neither is perfect. In the demo, 4 of 11 timestamps point at a line that doesn't state the claim, and the check kept a sentence that merged two facts.

Near it. In the demo, none of the 11 landed on a line that states the whole claim: 4 point at a line that doesn't state it, and the other 7 at a line that states only part of it. The 4 misses land early, and the supporting line follows within about 30 seconds. The overview at the top never gets a timestamp, by design, and a sentence the app can't place gets none rather than a guess.

It found all three voices in the demo and gave 5 of the 56 lines to the wrong one. No voice is stored or matched from one meeting to the next, so you name the voices once per meeting.

One meeting has been timed on the current build: 3 minutes 53 seconds of audio took about 6½ to 7 minutes, on an M4 with 24 GB, with the speech models already downloaded. The app sets no length limit, but I haven't timed a long meeting on this build, so I won't give you a per-hour figure yet.

Yes. Each one has an owner, and it gets a date only when the meeting gave one certain day. It's the part I'd check hardest. In the demo it listed 8 items for 5 commitments, with Dana's locum task in three wordings.

A tab where you put a question to a meeting in plain words. Each sentence of the answer comes with the transcript lines it cites. You can ask one meeting, or all of them at once.

It says “I found nothing in this meeting that says that.” and stops there. In the demo it did that for all 8 questions the meeting doesn't answer, or answers only in part, including one where a plausible day rate was one division away.

Yes. In the demo it refused one question the meeting did answer, and it added “dollars” to another answer when nobody named a currency. Asked again, it gave the same answer word for word, mistake included.

A second, different model has read the summary against the transcript and moved what it couldn't support into a separate card. This build shows one Checked mark for the whole summary. It doesn't mark sentences one by one.

Traced means a sentence carries a timestamp, placed by plain code with no model involved, on every setup. Checked means a second model read the summary and set aside what it couldn't support, and that needs Ollama. A timestamp can still point at the wrong line, and a checked summary can still hold a mistake.

No. In the demo it kept a sentence that merged two facts, and it passed an Ask answer that added “dollars”. It also sets aside sentences that are nearly right. It catches some mistakes, and reading the transcript catches more.

They go in a folded card under the summary, headed with a count, such as “4 unverified sentences”. The card's own line is “Read it as a lead, not a finding.” In this build they carry no timestamp.

It tells you why: the check model isn't downloaded, Ollama wasn't running, the check is off, or it stopped part-way. Check Now checks the summary without writing it again.

A Mac with Apple Silicon, running macOS 27 or later, plus Ollama. Ollama is a free app, and the setup assistant finds it or installs it for you. Older versions of macOS won't run it, and I haven't settled whether they ever could.

No. It's built for Apple Silicon only.

No. Meringo Minutes runs on the Mac only.

The app is 84 MB, with no models inside it. The default setup adds about 710 MB of speech models, Ollama, and two language models of 9.0 GB and 4.7 GB, so plan on about 15 GB of disk.

I don't know the minimum yet. I've measured it only on an M4 with 24 GB.

Ollama is a free app that runs language models, by default on your own Mac. Meringo Minutes uses it for the summary, the check and Ask. Without it, a meeting is still transcribed and its summary waits for Ollama. Apple Intelligence is listed in Settings as experimental.

Yes. Import Audio takes a file, and the demo on this site was made that way. It accepts m4a, wav, mp3, aiff and mp4.

$79 during the first month after launch, then $99. You pay once, and there's no subscription.

Not yet, and I don't have a date to give you. Email support@meringo.app and I'll send one note when it's ready.

One is planned, not built. The free version will process the first 30 minutes of each meeting, and buying processes the rest from the audio it already kept. There's no clock on it.

No. It'll be sold directly, through a link on this site.

No.

Yes, by default. As recording starts, the app says out loud: “I'm recording this meeting to take notes. The recording stays on my Mac.” You can change the words or switch it off, and it stays on until you do. A recording made without it is marked Unannounced.

No. Nothing joins the call and no calendar invite goes out; it records the call's audio on your Mac. That's why the announcement is on by default: it's how the others find out.

No, and it doesn't try. The reason is in the code: “the strictest participant's law governs, and nobody's location is knowable.” That's a design choice, not legal advice. One more limit: on headphones, the far end of a call may not hear the announcement.

Only if you turn that on. It's off by default, it works only while announcements are on, and it announces before it records.

The default speech model lists English and 24 other European languages, and Whisper, in Settings, lists 99. I've measured accuracy in English only: every recording in my benchmark is English, so I can't tell you how it does in the others. The vocabulary list works in English only.

The summary (with or without timestamps), the transcript and the mindmap, as Markdown, plain text, PDF, Mermaid or OPML, or everything to a folder at once. An exported summary says whether it was checked.

Yes. Add a word to the vocabulary list, and the app changes it wherever it also hears it in the audio. On three bench meetings full of uncommon names, read by synthesised voices, lists halved the error rate, from 10.9% to 5.4%. In the demo, “Pemberton” was fixed both times, and “locum”, heard as “locker”, stayed wrong.

Sometimes. When it notices a near-miss it offers the correction, and you confirm it. On the bench it made 14 offers, and all 14 were right. In the demo it made none.

It records your microphone and the audio your Mac plays, so the call app doesn't matter. It doesn't connect to any of them.

No. There's no sync and no account. You can export a meeting and send it yourself.

If a transcript is all you need, you may not need Meringo Minutes. The price is for the parts that help you check the notes: the timestamps, the second model's check, and an Ask that says when the meeting doesn't answer.

I do: Adam Siemaszko, under Meringo Labs LLC, in Erie County, New York. Write to me at support@meringo.app.