Documentation

Which local backend to auto-start: "mlx" (Apple MLX) or "ollama". Both speak the same OpenAI-compatible API; this only selects the server memo launches (see auto_start_server).

OpenAI-compatible base URL of your local model server. MLX: http://localhost:8080/v1. Ollama: http://localhost:11434/v1.

Model name the server should use. MLX: an mlx-community repo. Ollama: a pulled model tag, e.g. qwen2.5:0.5b (run ollama pull qwen2.5:0.5b first).

API key. Usually empty for a local MLX or Ollama server.

Where notes are stored. Leave empty to use the platform default (~/.local/share/memo/notes). Supports a leading ~ for your home directory.

Sampling temperature used when refining a note.

Maximum tokens the model may generate for a refinement.

How long (in seconds) to wait for the model to respond.

Stop sequences. Bigger instruct models sometimes keep generating past the end of their answer and leak chat-template markers. These sequences stop generation and are stripped from the output so they never reach your notes.

When true, memo starts the server selected by provider for you on launch and shuts it down when you quit. If a server is already listening on that port, memo leaves it alone, so for Ollama you can keep this on even when the Ollama app is already running.

When false, the server starts at launch and stays resident until you quit, so the model is always ready. Set to true to instead start it lazily on your first refine (Ctrl+R) and shut it back down once it has been idle, keeping the model out of memory until you actually use refinement. Only applies when auto_start_server = true.

On-demand mode only: how long to keep the server running after your last refine before shutting it down. 0 stops it immediately after each refine; a larger value keeps the model warm so repeated refines don't pay the reload cost.

Where to launch the server from. Point at your local mlx-lm checkout if you run from source; leave empty to launch from the current directory.

Virtualenv to run the server with. Either an absolute path / ~/path, or a name relative to mlx_repo (e.g. .venv). Empty uses python3 from PATH.

Show the keyboard-shortcuts hint bar along the bottom of the screen.

The key that drives dictation in the editor: hold for push-to-talk, double-press to toggle continuous live listening. A function key like "F5" / "f6", or a combo like "ctrl+k" / "ctrl+space".

Which whisper.cpp model to use. The matching ggml-<model>.bin is downloaded on first use into ~/.local/share/memo/models/. Options: tiny.en | base.en | small.en | medium.en | large-v3 (drop .en for multilingual). Bigger = more accurate but slower and larger.

Explicit path to a GGML model file. Leave empty to use the managed cache above. Supports a leading ~.

Spoken-language hint. Use a code like "en", or "auto" to let whisper detect it.

Live listening: how much trailing silence (milliseconds) ends a phrase and triggers its transcription.

Opens the command palette in the editor from anywhere in the line. Typing / at the start of a line opens the same palette. Same syntax as dictation_key: a function key like "F6", or a combo like "ctrl+space".

Toggles [ ] / [x] on the current to-do line.

The single prompt used when you press Ctrl+R. The default instructs the model to fix grammar and phrasing, preserve meaning and voice exactly, organize for readability with Markdown structure, and output only the finished note. Customize it to change how refinement behaves: make it translate, summarize, or match your house style.