How to read this demo
Unlearn asks a few adaptive questions, then walks a FalkorDB graph of prerequisites back to the earlier skill that actually broke, and remembers it next session. Every claim it makes links to the Cypher query behind it, under Receipts.
- Pick class
demo_video, choose Asha and click Start session. Answer, or let the simulated student answer, and watch the root cause light up amber in the graph. - Start a second session for Asha. The "Last time you…" banner is her memory, read back from the graph.
- Open Teacher for cohorts (
algo.WCCover shared misconceptions) and approve a proposal. Then Eval: graph against vector RAG against no memory.
Explain your thinking optional
- misconception
- mastered
- root cause
- asked now
- other
Boxes are skills (constructs). Solid arrows are REQUIRES edges, from a skill to the prerequisite it needs; the amber path is the chain from the topic down to the root cause.
Red pills are misconceptions this student showed (SHOWED edges in the graph), beside the skill where they weigh most; filled red means confidence above 0.6. Dashed red lines link a misconception to other skills it also appears in (SEEN_IN).
Green is mastered, amber with a thick border is the diagnosed root cause, the blue ring is the skill being asked about now, and grey is everything else in the slice.
Class demo
30 simulated students
What the Teacher agent suggests after a Tutor escalates a stuck student. Each one is a Proposal node in the graph; approving applies it from the graph's own edges and logs an event.
Students who hold the same confident misconception: links built from their SHOWED edges, grouped into connected components. A score alone can't show who is wrong in the same way.
The skills most of the topic depends on, ranked by PageRank over the prerequisite edges. A gap in one of these affects the most later skills.
Skills where misconceptions cross from one topic to another, by betweenness over the whole Eedi curriculum graph. A path question, so it needs the graph.
Evaluation
The same simulated students in every column: Unlearn traversing the graph (C) against no memory (A), vector RAG (B) and the whole graph pasted into the prompt (B+), all three on the same local model. Root-cause accuracy is the headline: finding the earlier skill that broke takes walking the prerequisites, not retrieving similar text.
Read live from eval/results/*.json: eval/run_eval.py, eval/concurrency.py and eval/latency.py write them.