From a passage to a practice attempt.
A 2 minute 50 second walkthrough of the working prototype, followed by its local architecture and measured limits.
Actual captured product screens with synthetic Kokoro narration. Development and demo production were assisted by Codex. Results use ten authored questions over twelve example passages; this is not an independent benchmark and no learning outcomes have been measured.
Read the narration transcript
RecallRoom is a study companion for students who want to understand and remember their own lesson notes. A fluent answer from an AI is not enough if you cannot inspect where it came from. RecallRoom starts with your text, keeps the original passages visible, and connects finding information with practicing it. Here is the working prototype, using an original plant biology lesson.
Ask, how do plants turn sunlight into food? In word search, exact vocabulary matters. Enabling local AI downloads a quantized MiniLM model and computes sentence embeddings on the device. Now the question can match photosynthesis even without naming it. The result is an exact excerpt, linked to source one. The similarity score describes how closely the meanings match. It is not a confidence score, and the full source remains available to check.
Next, practice the idea before revealing it. RecallRoom hides a phrase from a real sentence and creates options from terms in the same notes. This card asks for photosynthesis. After an answer, the app reveals the original sentence, highlights the correct phrase, and records whether the attempt was remembered or missed. You can also reveal the passage and self-check. The practice text is extracted from the source, rather than invented by a language model.
The review plan makes the next study decision concrete. It lists every passage and tracks recall attempts. Missed answers return sooner, while successful streaks move the next review further out. Unpracticed passages remain in the queue. These counts are a record of practice, not a diagnosis of mastery. The aim is a small, useful next step: return to the idea that needs another attempt, and keep its original context close.
The app has no account requirement or API key. Study notes, questions, embeddings, and practice records are processed in the browser. Only the public model, code library, and fonts are downloaded. Saving a room is optional and local to this device. Resetting clears the saved room. Students can export a source-linked study pack, enlarge text, or use an available on-device reading voice. This prototype does not synchronize notes to a cloud account.
On ten authored questions over the twelve-passage example, semantic search found the intended source first in seven cases and within the first three in all ten. Word search found it first in six. This is a small demonstration set, not an independent benchmark. English works best; ambiguous questions and awkward practice cards still need review. There are no claims of improved grades or measured learning outcomes. The project was built during LovHack with disclosed AI-assisted development. RecallRoom keeps the next study step connected to evidence you can inspect.