MemoryVerse AI '26

Setup: pip install -r requirements.txt, set ANTHROPIC_API_KEY, then streamlit run app.py This prototype demonstrates the full pipeline end-to-end: upload a document → AI classifies it (category, skills, dates, relationships) → browse by category → view a generated timeline → search in natural language (e.g. "show my AI projects") and retrieve the original file. Design decisions: classification and relationship extraction are done via LLM prompting rather than a trained classifier, given the timebox. Metadata is stored in a local JSON file instead of a vector database — the retrieval logic (natural language in, correct file out) is the same principle a vector-store-backed RAG system uses, just without an embedding index. Full reasoning and trade-offs are in the attached thought process sheet.