pi-learn is a project-local adaptive tutor for Pi. It builds a small learning plan, teaches one concept at a time, checks understanding with a quiz, and saves evidence so the next session can continue.
Chat history is not the learner model. Progress lives in files in this repository.
- Node.js 22 or newer
- Pi 0.84 or newer (
@earendil-works/pi-coding-agent) pi-subagents0.50 or newerpi-web-access0.23 or newer
Install Pi and the extensions, then start Pi from this repository:
npm install -g @earendil-works/pi-coding-agent
pi install npm:pi-subagents
pi install npm:pi-web-access
git clone https://github.com/arnavprabhu/pi-learn.git
cd pi-learn
npm install
pi --approve--approve trusts this project so Pi can load its local skills, extensions, agents, and settings. Pi may prompt for this approval interactively instead.
pi-learn works with every model provider Pi supports. Sign in through Pi, or run locally with Ollama. A 12B-class model or larger is enough to get started. With Ollama, that path is free and stays on your machine.
/teach Riemann sums
/frontier
/teach with no topic resumes the current mission. You can also ask in plain language: Teach me how Riemann sums approximate area.
/frontier prints the next concept from saved files, not from chat. When a mission is active, the footer shows the goal and the next concept id.
If /teach asks you to wait, let the current turn finish first.
Checks appear as a quiz overlay, not as “does that make sense?” Multiple choice is a list; free response is a short editor. I don't know is always an option and is better than guessing. The answer key is never shown in the UI.
Put readings, notes, code, or textbooks in knowledge/, then /teach again. Pi indexes changed files, searches relevant passages, and keeps a local cache. An empty folder is fine.
PDF, Markdown, text, HTML, JSON, CSV, and common source-code files are supported. PDF extraction is local and does not use OCR. knowledge/README.md is a guide for you and is not indexed as a source.
Files in knowledge/ are tracked by Git by default. Uncomment the knowledge rules in .gitignore if the material should stay local.
Probe → Plan → Teach → Verify → Persist
The tutor finds the nearest unmastered concept (the frontier), teaches that one step, then quizzes. “I get it” is not treated as evidence. A later session reads the same files and continues there.
| File | Role |
|---|---|
MISSION.md |
Current goal and depth cap |
learner/evidence.jsonl |
Canonical evidence log |
learner/concepts.json |
Derived mastery cache |
learning-records/ |
Compact daily notes |
knowledge/ |
Your source material |
.pi/knowledge-cache/ |
Generated local index |
Optional steering: learner/profile.md and learner/preferences.md. The tutor also writes MISSION.md when a topic starts.
Resets are never automatic. They ask for confirmation and keep knowledge/, its cache, and learning records.
/teach-reset topic <id>
/teach-reset mission
/teach-reset all
CLI equivalents:
npm run reset -- --topic riemann-sums
npm run reset -- --mission
npm run reset -- --all --yesSkills, extensions, agents, sessions, and learner state live in this repository. The project does not install anything into ~/.pi/ or use global Pi sessions. Pi itself may update its trust file if you approve the project. The researcher uses pi-subagents and pi-web-access; the tutor does not browse directly.
npm testTests use Node's built-in test runner and TypeScript's native type stripping.
.pi/ project-local Pi configuration and tutoring logic
knowledge/ learner-provided source material
lib/ inspectable mastery, graph, grading, and persistence code
learner/ saved learner state
learning-records/ session memory
scripts/reset.mjs explicit learner-state reset CLI
tests/ automated tests
MISSION.md current learning goal