Knowledge
Dungeon

Turn your notes into a dungeon. Every chapter is a room, every room has a boss, and an AI Dungeon Master sends back the ones you're forgetting.

✦ trained on 325k real student answers ✦ tracer AUC 0.759 on unseen topics ✦ bosses write questions from your own notes ✦ every question double-checked by a blind verifier ✦ forgotten bosses respawn ✦ spaced repetition 1 → 3 → 7 → 14 → 30 days ✦ trained on 325k real student answers ✦ tracer AUC 0.759 on unseen topics ✦ bosses write questions from your own notes ✦ every question double-checked by a blind verifier ✦ forgotten bosses respawn ✦ spaced repetition 1 → 3 → 7 → 14 → 30 days

How it works

01_map.exe

Map

An LLM reads your notes, or a whole textbook's chapter list, and carves one room per topic, locked behind the topics it builds on.

02_fight.exe

Fight

Bosses ask questions written from your notes. The weaker the topic, the tougher the boss and the longer the fight.

03_remember.exe

Remember

A knowledge-tracing model predicts what you're forgetting. The Dungeon Master respawns those bosses and points you to the next room.

The model

Classic knowledge tracing learns a fixed list of skills, so a model trained on maths can't judge your DBMS notes. Our tracer learns how people learn instead: attempts, streaks, recency and accuracy, signals that exist for any topic.

It trades some accuracy for working on topics it has never seen, which is exactly what a game built from your own notes needs. Trained and tested on ASSISTments 2009 (325k answers).

tracer_eval.log

Overall accuracy (baseline)0.682

Logistic regression0.725

Our topic-agnostic tracer0.759

DKT LSTM (fixed skills only)0.819

Test AUC, higher is better

KnowledgeDungeon

How it works

Map

Fight

Remember

The model

Knowledge
Dungeon