What Lumen is

Lumen brings your recordings, files, and notes into one app and processes them locally. The goal is a personal agent that helps you turn what was said and written into finished work, without sending any of it to a server.

What it does today

Meetings. Record your microphone and the call, or import an existing recording. Lumen transcribes it, catches commitments as you talk, and writes a summary with decisions and next steps, each linked to the transcript. Before a meeting it shows a brief of what you agreed last time, and afterwards it drafts a follow-up for you to send. Action items go to Reminders when you approve them.

Search and Ask. Press ⌘K to search the folders you choose and your meetings by name, number, or meaning, including scanned PDFs, images, audio, and video. Switch to Ask for an answer with its sources. Results open at the matching page or moment. See the search guide and the meetings guide.

Memory and Today. Turn on memory sources one at a time: calendar, meetings, chosen folders, app activity, and clipboard. Lumen suggests things to remember, and you decide what it keeps. With app activity on, Today shows your time by app, category, and project. Every source is off by default. See the memory guide.

Routines. Small automations, such as tidying Downloads or filing meeting notes, that show every change for you to review before anything moves. Files go to the Trash, never permanently deleted, and you can undo. See the routines guide.

How the local AI works

Lumen uses Whisper large-v3-turbo for transcription, Gemma 4 E2B for notes and answers, and EmbeddingGemma 2 for search and memory. When you choose Get started, Lumen downloads them once from Hugging Face, about 6 GB, and checks every file against a pinned checksum. After that they run on the device. Everything else Lumen needs, including its Python runtime, ships inside the app, so there's nothing else to install. Lumen never falls back to a cloud service. Keyword search and the sample meeting work before the download finishes. See How local AI is prepared.

Privacy and its limits

Lumen has no account and no telemetry. Only a small download helper can reach the network, and it only fetches the models from Hugging Face; the app and its AI processes have no network access. Lumen reads the folders you choose. It records when you start a recording, or when you turn on automatic recording, which is off by default. Memory is off until you turn on a source, and you can pause it, exclude apps, or delete history at any time.

Lumen stores your data in local storage and doesn't add its own encryption, so disk protection comes from your operating system, such as FileVault on macOS. Lumen hasn't had an independent security audit. For details, see Privacy and permissions. This website is separate from the app; website privacy covers it.

What's been tested

Development and testing have used one M1 Pro with 16 GB of memory. Transcription, generated notes, and search have passed small evaluations with synthetic inputs, and one saved recording has been processed end to end. That's evidence the pieces work, not a measure of accuracy across languages or of support for other hardware. Live call capture still needs testing across devices. The preview status page tracks what's done and what's still open.

Windows and Linux versions are planned, and each will run its models on that device. Mobile, sync between devices, and dictation aren't available yet. Memory, Today, routines, and the meeting copilot are new and have been checked with tests and synthetic data; real-world testing is under way.

Get in touch

The source repository isn't public yet. To get on the list, request early access and tell us which platform you use. For anything else, email us. A description is enough; there's no need to send recordings or documents. Emailing doesn't sign you up for a mailing list.