Setup guide

Get Retain running
in under 10 minutes

AI provider, transcription, speaker identification, and calendar sync — all covered here.

1. Choose your AI provider

Retain needs an AI model to generate summaries, titles, and chat responses. You can switch at any time in Settings → AI Provider.

Not sure which to pick?
→ Claude if you want the best summaries and don't mind a small per-use cost→ Ollama if you want everything on-device with no account or API key

Claude (Anthropic)

Recommended

Best quality summaries. Handles nuanced questions exceptionally well. New accounts get free credits.

Ollama (local)

Fully private. No API key, no subscription, no data leaving your Mac. Needs ~8 GB RAM.

Option A — Claude by Anthropic

Go to console.anthropic.com and sign in or create a free account.

Click API Keys in the left sidebar, then Create Key.

Copy the key — it starts with sk-ant-…

In Retain, open Settings → AI Provider.

Select Claude (Anthropic) and paste your key.

Click Save. Retain verifies the key immediately.

A typical one-hour meeting summary costs less than $0.01 with Claude. New accounts receive free credits.

Option B — Ollama (fully local)

Download and install Ollama — it runs as a background service.

Open Terminal and pull a model:

bash

# Recommended — best quality
$ ollama pull gemma3:12b
# Faster on older hardware
$ ollama pull qwen3:8b
# Lightweight (< 4 GB RAM)
$ ollama pull gemma3:4b

Hardware note: gemma3:12b requires ~10 GB RAM. Use gemma3:4b or qwen3:8b on 8 GB Macs.

In Retain, open Settings → AI Provider, select Ollama (local), and choose your model from the dropdown. Retain detects running models automatically.

2. Transcription

Retain transcribes audio locally using Whisper by default — your recordings never leave your Mac. No setup required. The first time you finish a recording, Retain downloads the Whisper model in the background (~145 MB for base).

Choose your model in Settings → Transcription:

Model	Size	Speed	Accuracy
tiny	75 MB	Fastest	Good for short calls
basedefault	145 MB	Fast	Good
small	465 MB	Moderate	Better
medium	1.5 GB	Slower	Great
large	3 GB	Slowest	Best

Start with base. Upgrade to small or medium if you notice errors with accents or technical terminology.

3. Speaker identification

Speaker identification labels each transcript line with who was speaking. It runs after each recording using Pyannote, an open-source library.

Without speaker ID

"Ok so the budget is
looking tight for Q3
and we should probably
revisit the roadmap…"

"Agreed. Let's look at
what we can cut."

With speaker ID

Speaker 1:
"Ok so the budget is
looking tight for Q3…"

Speaker 2:
"Agreed. Let's look at
what we can cut."

This feature requires two one-time setup steps.

Step 1 — Install Pyannote

bash

$ pip3 install pyannote.audio

If you see pip3: command not found, install Python 3 via python.org or brew install python.

Step 2 — Accept the model licence on Hugging Face

Create a free account at huggingface.co.

Go to pyannote/speaker-diarization-3.1 and click Agree and access repository.

Go to Settings → Tokens → New token. Set access to Read and copy the key (starts with hf_…).

In Retain, open Settings → Transcription, paste your token, and toggle Enable speaker identification on.

On the first run Pyannote downloads the model (~1.3 GB) — allow a few minutes.

4. Calendar sync

Retain connects to your calendar to show upcoming meetings and link recordings to events. Calendar data never leaves your Mac.

macOS Calendar (simplest)

Open System Settings → Internet Accounts → Add Account and sign into your Google account.

Make sure Calendars is toggled on, then open the macOS Calendar app to confirm your events are visible.

In Retain, open Settings → Calendar and click Connect Calendar.

macOS will prompt for Calendar access — click Allow.

Google Calendar direct (advanced)

Use this only if your organisation restricts the macOS Calendar app. Requires a free Google Cloud project.

Go to console.cloud.google.com, create a project, and enable the Google Calendar API.

Set up an OAuth consent screen (External, read-only calendar scope).

Create an OAuth Client ID (Desktop app) and copy the Client ID.

In Retain, open Settings → Calendar → Advanced → Google Client ID, paste it, and click Connect Google Calendar.

Troubleshooting

Check your AI provider key in Settings → AI Provider. If you're using Ollama, make sure the Ollama app is running — look for it in your menu bar or run ollama serve in Terminal.

Switch to a smaller Whisper model in Settings → Transcription. tiny or base are significantly faster than medium or large.

Confirm Pyannote is installed (pip3 show pyannote.audio in Terminal) and that your Hugging Face token is entered in Settings. The model takes a few minutes to download on first run.

Open macOS Calendar and confirm your events appear there first. Then go to System Settings → Privacy & Security → Calendars and make sure Retain has access.

Retain detects meetings via process name. Make sure Zoom or Teams is fully open (not just in the dock). If auto-detect still doesn't trigger, you can start recording manually from the menu bar.

Ready to record your first meeting?

Download for Mac, Free

Get Retain runningin under 10 minutes