RecallWeave

Offline adaptive learning with transparent Bayesian models, reusable courses, lesson authoring, saved progress and interactive labs. Learn connections, then explain and apply them.

Comment

Inspiration and impact

RecallWeave helps people learn the connections between ideas, then explain and apply them. Built for ForgeHacks AI + Education, it began with six original cellular-energy questions and has grown into an offline learning toolkit: reusable lesson decks, a course catalog, authoring tools and interactive subject labs.

Try it now

These are working browser pages, with no installation, account or API key required. The standalone HTML files also open directly from disk.

What it does now

Adaptive learning with visible reasoning. A local Bayesian Knowledge Tracing model updates estimated concept knowledge using explicit initial-knowledge, learning, guess and slip probabilities. Expected information gain plus a small prerequisite-repair bonus chooses the next unanswered question. Learners receive authored explanations and transfer prompts. Estimates are visibly labeled as model state, not grades.

Review, practice and reflection. The completed learning trace preserves every first answer and explanation. Missed-question practice records retries separately without rewriting the initial answers or mastery estimates. Learners can write their own explanations and application responses, then download readable study notes.

Portable progress. Explicit JSON downloads let learners save and resume an unfinished lesson or restore a completed trace against the same course and model. A separate comparison page reads two compatible saved traces side by side; it does not treat answer differences as proof of learning improvement.

Reusable teaching tools. Bring-your-own-deck import checks structure, answer identities and prerequisite cycles before an explicit preview/start step. Deck studio creates and reopens drafts, validates lessons and exports checked decks. Lesson focus retains selected concepts and their prerequisite closure; the handout builder produces separate learner worksheets and teacher answer keys. The live catalog currently lists 11 courses, while the repository contains additional original lessons and standalone labs across biology, mathematics, physics and computing.

Interactive explorations. Examples include RC transients, coupled motion, Fourier reconstruction, Mendelian crosses, binary search, recursion, network flow and numerical precision. Labs expose applied inputs, intermediate steps or physical quantities, assumptions and downloadable observations. Their course files can be studied through the same learner.

How we built it and what we learned

HTML, CSS and modular JavaScript implement the learner, transparent probability model, deck validation, authoring and subject explorers. Python and Node build scripts generate self-contained offline HTML and the course pack. The project was substantially created during this hackathon starting October 5; it is not a recycled contest entry.

The main engineering challenge was preserving meaningful learning records while adding import, practice, reflection and restore flows. Canonical answer identities survive shuffled display choices; retries stay separate from first answers; invalid or stale file previews cannot silently replace an active lesson. The current source passes all 739 Node unit/property tests (October 10, 2026). The repository also includes generated-file checks and browser acceptance evidence for the learner and newer tools. The screenshots show the current deployed learner, catalog and circuit lab.

Content, privacy and limitations

The bundled biology questions use original wording informed by OpenStax Biology 2e; current source references and permissions are recorded in the deck and content-review documentation. Other courses include their own attribution and permission statements.

Learner files are read locally. Answers and writing stay in the current tab unless explicitly downloaded; there is no automatic browser storage or learner-data upload. Reloading clears unsaved state. The hosted site delivers static files; the learning logic makes no hosted AI calls.

The scripted learner comparison is synthetic, not an efficacy study. Model parameters are illustrative defaults, not fitted or validated assessments. Structural deck validation does not establish subject accuracy or reuse rights. Next steps are educator review, parameter calibration and consent-based evaluation of learning outcomes.

AI assistance disclosure

AI assistance helped develop and test the implementation and draft original educational wording. Runtime adaptation is local Bayesian Knowledge Tracing and deterministic selection, without an LLM or provider API. Human subject-matter review remains necessary before educational deployment.

Built With

Share this project:

Updates

Submission history