MedMemGraph

Context

Built for HackHydra — Track 03 (Memory & Context Retrieval). A clinical assistant that spans many hospital admissions needs real memory: it has to recall what happened last visit, compare across visits, and — critically in a medical setting — never make something up.

Problem

The default pattern, chunk-similarity RAG, ranks passages by embedding distance. That loses provenance (you can't trace an answer back to the fact that justified it) and it fails silently on absence — ask about something the record doesn't contain and a similarity search still returns the "closest" chunk, inviting a confident, wrong answer.

Approach

• Facts as versioned graph nodes. Multi-admission conversation histories are parsed into clinical facts stored as versioned nodes in a property graph, rather than flattened into text chunks.
• Retrieval by traversal, not ranking. Questions are answered by walking weighted paths through the graph (HydraDB SPpaths / SSpaths / MSpaths path-enumeration), so a "why do you believe this?" query returns the actual provenance trail it traversed.
• Structural absence detection. When no supporting path exists, the system declines to answer instead of guessing — turning "I don't have that" into a first-class, structural outcome.
• Cross-admission reasoning. Multi-hop walks synthesize comparisons across visits ("how has this changed since last admission?") that a single chunk lookup can't express.

Results & what I learned

• A working end-to-end prototype: ingest histories → build the fact graph → answer questions with a provenance path attached and an explicit refusal path for unanswerable ones.
• Learned: for high-stakes retrieval, representing knowledge as a traversable graph buys you two things vector similarity can't — explainability (the path is the citation) and honest absence (no path ⇒ no answer).