Inspiration
Most memory tools today rely on rigid flashcards or repetitive drilling. They don’t reflect how the human brain actually processes and recalls information. We were inspired to build Recap to simulate the brain’s layered memory structure and dynamic forgetting behavior, combining spaced repetition with semantic evaluation to deliver smarter, more human learning experiences.
What it does
Recap is an AI-powered memory planning system that helps users remember knowledge more effectively across languages, sciences, and beyond. It builds a multi-layer knowledge graph (e.g., sentence → word → morpheme → letter for English, or unit → topic → concept → core principle for physics), and tracks how each knowledge point decays over time. The system then chooses what to review next by maximizing the number of related nodes that can be reinforced in one session.
In language learning, users chat naturally with an AI. The AI brings up vocabulary they are close to forgetting and evaluates their understanding. In STEM subjects, Recap embeds review into problem-solving workflows. Users can also upload documents or URLs to automatically convert reading material into a personalized memory map with live decay tracking and embedded quizzes.
How I built it
Recap has both a frontend and backend built from scratch during the hackathon. The backend is built using FastAPI with real-time retention updates, knowledge graph traversal, and a spaced repetition model based on an exponential forgetting curve. The frontend is a React-based interface with dynamic visualizations of knowledge graphs and AI-guided review workflows.
We also integrated a multi-agent system using the TEN Framework in a separate demo, which coordinates language understanding, semantic evaluation, and memory state updates. The Google Maps Street View Static API was used to link visual exploration with vocabulary recall in conversational learning.
Challenges I ran into
One of the biggest challenges was designing a memory algorithm that could propagate retention changes through graph layers and select the most efficient review paths. Coordinating semantic scoring, decay modeling, and multi-agent dialogue logic was another major challenge, especially under hackathon time constraints.
Integrating document parsing, scheduling real-time decay updates, and maintaining a stable state across sessions also required careful design and iteration.
Accomplishments that I'm proud of
I built a fully working memory planning engine that supports multi-layer knowledge graphs and real-time adaptive review. The AI can guide users through custom content naturally, while optimizing memory reinforcement intelligently. I’m especially proud of the retention-aware selection algorithm that chooses batches of nodes to review for maximum efficiency.
What I learned
I learned how to combine cognitive science principles with AI-driven evaluation and scheduling logic. I also deepened my understanding of graph algorithms, agent collaboration, and system design under constraints. More broadly, I gained experience in building tools that align with how humans actually think and forget.
What's next for Recap
I plan to scale Recap into a full learning platform supporting multiple subjects and languages. Future versions will include user-editable knowledge graphs, long-term learning analytics, and more multimodal inputs. The ultimate goal is to provide students with a system that actually understands what they know and what they’re forgetting—making cramming and inefficient studying a thing of the past.
Log in or sign up for Devpost to join the conversation.