GitHub 45 Download
  1. Docs
  2. Students
  3. AI assistant

AI assistant

The experimental, fully local AI assistant; how to enable it and what it can do.

Mnemo includes an experimental AI assistant. It is disabled by default, runs entirely on your own machine, and is still under heavy development. Expect rough edges.

Enabling it

  1. Open Settings → AI Tools and turn on the assistant. A warning dialog explains the experimental status first.
  2. Open the model manager from the same page and download the local models. Mnemo detects your hardware and suggests a model tier; stronger machines get larger models.

Once enabled, a Chat entry appears in the sidebar and an assistant panel becomes available at the right edge of the window. Models load on the first message, so the first response after a cold start takes noticeably longer than later ones.

What it can do

  • Chat with streamed answers and conversation history. Response length is adjustable (short, normal, detailed), and you can choose between a fast model and a stronger reasoning model, or let Mnemo route automatically.
  • Work inside the app. The assistant can read, create, and edit notes, build and edit mind maps, look up your study statistics, change settings, and navigate to pages, when you ask it to.
  • See images. Attach an image to a message and the assistant can describe or answer questions about it, if the optional vision models are installed.
  • Take dictation. Voice input is transcribed locally with Whisper.
  • Remember context. Long conversations are summarized in the background so the assistant keeps track without rereading everything.

The assistant also powers learning paths.

What it cannot do yet

  • Generate flashcards from notes.
  • Read non-image file attachments in chat. You can attach them, but their contents are not passed to the model.

Hardware and performance

The models run through llama.cpp on your CPU, or on your GPU when GPU acceleration is enabled in the AI settings. Idle models are unloaded after a configurable timeout to free memory. Response quality and speed depend directly on your hardware and the chosen model tier.

Privacy

Everything stays on your machine. Prompts, answers, chat history, and the models themselves are local. The only network traffic the AI features create is the one-time model download from GitHub. There is no API key, no cloud account, and no telemetry.