learn-engine

AI-native Education Intelligence Platform — a personal university of research agents, teaching agents, and knowledge graphs, run against an evidence-based learner model. Fourth engine on the research-engine substrate. A MoreSalamander StudioLabs production.

Student Goal → Roadmap DAG → Parallel Educational Research
→ Student Context Graph → Education Knowledge Graph
→ Teaching Agents → Adaptive Learning Engine
→ Practice + Assessment → Mastery → Shared Teaching Memory

Mastery is computed, never asserted: lesson → introduced; first correct → practicing; mastered needs a three-streak spanning two item kinds; two misses demote. LLM-judged answers carry half weight. Review timing is a fixed spaced-repetition schedule. And generated code exercises pass the code-executes gate — the reference solution runs in a sandbox and must reproduce its answer, or the exercise never reaches the learner. When the model's claim disagrees with execution, execution wins.

The adaptive policy states its reasons

Every activity carries WHY it was chosen: overdue review (spaced repetition outranks everything) → misconception repair (Socratic re-teach of the exact confusion) → continue practicing → next concept in prerequisite order → done. Mastered concepts are never re-taught; cross-disciplinary bridges from what the learner already knows decorate new-concept activities.

What the graphs remember

Student Context Graph	per session: current activity + reason, confusion points, interactions, mastery snapshot
Education Knowledge Graph	persistent: concept relations across subjects, researched resources, the shared misconception library, and explanation-outcome records — which teaching method actually precedes success, per concept and per learner

Honest scope

Sources are the live open web (arXiv, OpenAlex, Semantic Scholar, Wikipedia, GitHub, HN, Crossref, GDELT; open courseware best-effort) — no LMS integrations. Mastery evidence comes only from graded interactions inside the system. Code practice is Python-only in an isolated subprocess. Visualizations are mermaid concept maps generated deterministically from the graph, bundled with each lesson as markdown + static HTML.

Run it

.venv/bin/python -m learn_engine study "Teach me machine learning" --steps 3
# or the API on :8020 — reference answers never ship to the client

github.com/MoreSalamander/learn-engine · 31 offline tests · substrate: research-engine · part of MoreSalamander