Most AI tutors are goldfish. They reset their memory every session, treating every query as a blank slate. LearnHarness, a new open-source agent harness by amosblomqvist, is trying to fix that by building a persistent knowledge graph of your actual understanding. The tool uses AI agents to plan lessons based on your current reasoning state, not just your search history, creating a dependency map from where you are to where you want to be.

The Measure-Plan-Teach-Practice Loop

The workflow is strictly structured. First, the agent 'Measures' your understanding by analyzing your answers, confidence levels, and thinking patterns to find exactly where your knowledge ends. It then 'Plans' a dependency map to your goal, which you can inspect and which adapts if your understanding surprises it. The 'Teach' phase builds on known concepts, forcing you to predict outcomes before getting explanations, ensuring confusions are repaired before moving on. Finally, 'Practice' sessions pose varied problems that require recalling knowledge without prompts, adapting hints to your reasoning style.

Your Keys, Your Data, Your Graph

LearnHarness is built for the self-hosted crowd. It runs locally on macOS, Windows, and Linux, storing chats, settings, and API keys directly on your machine. You bring your own models, supporting provider API keys, subscription connections, Claude Code CLI, or custom OpenAI-compatible endpoints like local Ollama servers. The knowledge graph stores propositions and reasoning chains as plain files, meaning your learning record is inspectable and portable. It’s licensed under AGPL-3.0, with a promise that future charges will only cover convenience features like managed keys, never core functionality.

Key Takeaways

  • Persistent state: Unlike standard chatbots, LearnHarness carries a knowledge graph across sessions, tracking reasoning chains rather than just ratings.
  • BYO-Model flexibility: Supports local models via Ollama, Claude Code CLI, and various API providers, with no vendor lock-in.
  • Active learning: Forces users to predict outcomes and explain reasoning, moving beyond passive consumption of explanations.
  • Open source: Released under AGPL-3.0 with visible, editable system prompts and local storage for all data.

The Bottom Line

This is the kind of agent architecture we’ve been waiting for: one that actually learns the user, not just the topic. If the knowledge graph implementation is as robust as the documentation suggests, this could finally make AI-assisted learning stick. Note: The source material does not include user testimonials or benchmark data on retention rates, so real-world efficacy remains to be proven by the community.