StudyBuddy: AI Study Assistant with Memory

Most AI study tools forget you the moment you close the tab. StudyBuddy doesn't. StudyBuddy is an AI study assistant built with genuinely persistent memory across two layers: a structured SQLite database tracks every concept a student studies and their mastery level, while a ChromaDB semantic vector store enables meaning-based recall, finding related notes even when wording differs. On top of this sits an SM-2 spaced-repetition engine modeling the Ebbinghaus forgetting curve, automatically calculating when each concept needs review before it is forgotten. The entire system runs through Fireworks AI, serving inference on AMD hardware. StudyBuddy uses confidence-based model routing: a lightweight classifier evaluates request complexity in real time, dispatching it to the right-sized model automatically, with a fallback ready. This was a deliberate engineering decision, I evaluated Gemma directly through Fireworks but found its on-demand hourly billing unsuitable for a live-tested hackathon app. I architected around fully serverless models to guarantee reliability for judges testing at any time. Beyond the API, I validated direct engagement with AMD GPU infrastructure: a standalone notebook (amd_gpu_proof.ipynb) runs local embedding generation using ROCm-based PyTorch directly on an AMD Developer Cloud GPU pod. This confirms hardware detection, real inference execution, and timing natively. An agentic orchestration core, built on native function-calling, decides autonomously when to save sessions, generate personalized quizzes from history, identify weak areas, or build study plans, six tools working seamlessly without explicit commands. A live dashboard visualizes mastery per concept, accuracy trends, and streaks from persisted data. StudyBuddy is free and open-source, built for students who cannot afford private tutoring, giving anyone a study companion that actually remembers them