Turning on AI

AI is off until you turn it on and give Myne something to run. This page walks through enabling the beta opt-in, choosing model files, what “ready” means, and how to keep chosen folders out of AI entirely.

Prerequisites

• An unlocked vault.
• Something for Myne to run: a model file you downloaded yourself, a model runtime already running on this computer, or a remote provider you name. Myne never downloads a model. See Choosing model files below and Where your AI runs.

Enabling local AI, on a computer

These steps are the computer’s. A phone has no Developer mode and no Developer group, and reaches AI from Settings directly; skip to Enabling AI on a phone below.

1. Open Settings from the gear in the bottom bar or with ⌘,.
2. Open the About panel and turn on Developer mode. The AI features live in a Developer group that stays hidden until you do this.
3. In the panel rail, open AI under the Developer group that now appears.
4. Turn on Enable local AI. The helper text says what the toggle does and does not settle: it runs a model you provide, whether your notes leave this device depends on the provider you name below, and the feature is experimental and off by default.

Below the toggle, one sentence never changes (Myne never sends your notes anywhere you have not named, and never downloads a model) followed by a headline stating the configuration you are actually in: a model on your device, a program on your computer, or a provider you named. Under those, a line reads Answering now: with a name, or Nothing is set up to answer yet. The How does this work? disclosure expands the honest detail, including the part Myne does not protect against. That copy is worth reading once; it is the same boundary described in AI privacy and limits.

Enabling AI on a phone

The steps above are the computer’s. A phone has no Developer group and no model files to choose, because no model on a phone answers questions or indexes your notes. (Transcription and image text do run on the phone, on models that ship with Myne; they are not part of this switch.)

1. Open Settings and tap AI. It is a top-level entry, not nested under another category.
2. Turn on Use AI on this phone. The helper says what that commits you to: “Off by default. No AI model on this phone answers questions or reads your notes for AI, so every request goes to a service you name below, and your notes leave this device when it does. Transcription and image text still run here, on models that ship with Myne.”
3. Choose a provider. Choosing one for the first time on this phone opens the consent screen described in Where your AI runs, and nothing is stored or sent until you accept it there.
4. Paste the provider’s API key, then load the model list and pick a model. Myne ships no default model and will not choose one for you.

A phone you add to an account that already has AI set up is not finished being set up. The provider, the model and the key travel with your vault; the approval does not. That phone shows the provider as chosen and a row reading Approve sending on this phone above the key field. Until you open that and accept, the phone sends nothing, and asking anything says exactly that rather than failing vaguely.

The reason is worth one sentence: agreeing on your laptop is a statement about your laptop. A phone is a different device, in a different pocket, on a different network, and the screen that says what leaves it should be read on it.

Suggesting tags automatically

Under the toggle sits a second switch, Suggest tags automatically, off by default. It has no model file to choose: tag suggestion runs on a model that ships inside Myne. If your build ships without that model, the switch is off and cannot be turned on, and Myne says so: it never downloads one. See Suggest tags.

Choosing model files

Myne lists three model roles under Models in the same panel. You bring the first two; the third is an override.

• Text model: chat, summaries, writing. Used for chat answers, summaries and writing help. The recommended starting point is Qwen2.5-1.5B-Instruct (GGUF). This role needs the .gguf file plus tokenizer.json in the same folder.
• Search model: semantic search & Q&A. Used to index your notes so questions can find the right passages. The recommended starting point is all-MiniLM-L6-v2. This role needs model.safetensors, config.json, and tokenizer.json in one folder.
• Speech-to-text model (transcription). Optional. Myne already ships a speech-to-text model, so this picker is only for replacing it with one of your own. The recommended starting point is whisper-base (multilingual), and the role loads GGUF files. Speech-to-text has its own switch in Settings → Transcription, is on by default, and is not part of the AI beta. See Transcribing a voice note.

Each role shows a recommended-model name as a link. Clicking it opens the page in your own browser as an operating-system handoff; Myne does not fetch anything. Download the files there, then:

1. Next to a role, click Choose file….
2. Pick the file in the picker (the .gguf for the text model, the .safetensors for the search model).

The chosen path is stored on your device only and the file is never copied into your vault. The path must be outside your vault folder; if you pick a file inside the vault tree, Myne refuses it and tells you it must be outside your vault.

Model files are one of three ways to give Myne something to run. You can instead point Myne at a model runtime already running on this computer, under Local model runtime, or name a remote provider under Remote provider. Both are covered in Where your AI runs, including what leaves your device in each case.

What “ready” means

Enabling the toggle is not enough on its own. AI is ready only once something is set up to answer: a text model file, a local runtime, or a provider. The AI Chat tab’s header shows which state you are in:

• No provider: AI is on, but nothing is set up to answer. Choose a model file or name a provider in Settings.
• Not answering: what you set up is not responding. Check that it is running, then try again.
• Unavailable: this build of Myne was compiled without local AI, so nothing here can run.
• No search model: something can answer, but no model on this device can search your notes. You can still send a message and a run still starts against the notes you gave it; what stops working is the run’s ability to look beyond that scope. Choose a search model in Settings → Developer → AI.
• Preparing…: indexing your notes for the first time in this session (see below).
• Ready, or the provider’s name: set up and working. With a provider named, the header shows its name instead of “Ready”, so you can see at a glance where your questions are going.

With AI off there is no header to read: the AI Chat tab is not shown at all.

When you unlock the vault with AI on, Myne builds a search index of your notes on this device so that Ask your notes can find relevant passages. That index is encrypted at rest and reused across sessions, so after the first build it only reconciles what changed. It is never synced and never backed up, and it is never built by a provider.

Excluding folders

The Excluded folders box lets you name folders that AI must never read or index. Enter one folder path per line. A note is excluded when its folder is one you listed or is nested inside one, so excluding Private/ also excludes everything under it. The match ignores case, so a folder stays excluded regardless of how you capitalized the path.

Exclusion is enforced both when notes are indexed and when they are queried, and every AI surface checks it before the model is ever called. A note in an excluded folder is never embedded, never summarized, never used to ground a chat answer, and never offered tag suggestions or writing help.

Excluding a folder travels to your other devices, and un-excluding one does not travel at all. The exclusion list only ever grows when devices sync: a folder you exclude here shows up as excluded everywhere, which is what stops a second device from indexing something the first was told to leave alone. The same rule makes a removal hard to keep. On a vault that syncs, taking a folder off the list is undone by the next sync round, because the other side still carries the entry and the two lists are combined rather than compared. Doing it on every device does not help either: whichever device syncs first brings the entry back. So treat an exclusion as permanent for as long as the vault syncs, and if you need one undone, say so on the issue tracker rather than working around it. On a vault that has never been synced, a removal simply holds. The CLI exclusion list works the same way, for the same reason.

Myne can remember a large but limited number of folders across all of its folder lists together. If you reach that limit it says so when you add one, and deleting a folder you no longer need makes room.

What the assistant remembers

Agent runs can remember things about you between runs: facts they learned while working, such as how you like answers written. This memory is off the same switch as everything else on this page. A vault that never turned AI on carries no memory at all.

You do not approve each thing it remembers. Instead, everything it writes is visible and editable, in two places. The memory button in the AI Chat tab header opens the live list beside the conversation. Settings → AI holds the full panel: the same list with edit and delete, a Clear all memory button that removes every entry after a confirm, and a voice-preferences editor, one instruction per line, for standing rules about tone, length, or language. A run that wrote something says so on its own turn with a one-line marker, so the write is never silent.

Three properties are worth knowing. First, at most one entry per run: a run cannot fill your memory in one go. The roster holds at most 100 entries; when it is full, new entries are refused until you edit or delete one. Second, memory only learns from notes AI was allowed to read. The run that learns is the same run whose reading the exclusion box above constrains, so anything remembered came from notes AI could see. There is no second exclusion check at the moment of remembering; the boundary is the reading, not the writing. Third, memory entries and voice preferences are vault content like notes: encrypted at rest, synced to your other devices, and included in backups.

Limits

On a phone, “turning on AI” means turning on sending. There is no on-device model to fall back to, so with the switch on and a provider named, the text of the notes involved in each request leaves the device. Turning the switch off, or setting the provider back to none, stops that; the second also removes the key and this phone’s approval.

Turning on AI adds a model runtime to the Myne process. That process holds your decrypted notes and keys, and the model is not isolated from them; a model file you download is untrusted input that Myne does not vouch for. Turning the toggle on does not by itself send anything anywhere: that depends on what you name, and Where your AI runs is where that choice is made and explained. Excluding a folder keeps its notes out of AI reading and indexing, but it is a Myne-level control rather than an operating-system boundary, and it does not protect the folder against malware already running on your device, which can read app memory regardless. The encrypted index Myne builds for search is a rebuildable local cache; it is never uploaded. Full detail is in AI privacy and limits.

Shortcuts