amosblomqvist/learnharness

What it solves

Traditional lessons give the same explanation to everyone, ignoring what you already know or how you think. LearnHarness instead first measures your current understanding and reasoning on each part of a lesson, then builds a teaching plan tailored to you. It aims for deep understanding, where the goal feels obviously true, not just familiar.

How it works

1. Measure. The agent asks questions to find where your understanding ends, using your answers, confidence, and thinking.
2. Plan. It creates a dependency map from your starting point to your goal. The plan is visible and adapts if your understanding turns out different.
3. Teach. Lessons build on what you already understand: you predict, get explanations, and check understanding. Confusions are fixed before building on them.
4. Practice. Separate sessions pose varied problems that require you to recall the right knowledge without it being named first. Hints and problems adapt to your reasoning.

You can "think out loud" in text or via dictation, giving the agent a detailed picture of your reasoning. It can also research questions in the background, create visual explanations, and build interactive demonstrations.

A persistent knowledge graph stores what you learn as propositions and reasoning chains, with links back to source sessions. Future sessions read this graph instead of starting over, and practice sessions draw material from it.

Who it's for

People who want to learn a subject with a structured, reasoning-based approach and who value deep understanding over rote memory. It's also for those who care about data privacy, since everything is stored locally, and who appreciate open-source software with visible, editable system prompts.

Highlights

• Uses AI agents to measure your understanding and plan tailored lessons.
• Adaptive teaching and practice that respond to your reasoning.
• Persistent knowledge graph that carries learning across sessions.
• Local storage of your chats, graph, settings, and API keys.
• No subscription; you only pay your model providers directly.
• Cross-platform: macOS (Apple Silicon), Windows, and Linux.
• Open source under AGPL-3.0.