Show HN: Open-source RAG server with retrieval visualization (Postgres+pgvector) (github.com)

MemVault is an open-source, production-ready RAG (retrieval-augmented generation) server that packages long-term memory for AI agents into a single self-hosted API backed by PostgreSQL + pgvector. It automates chunking, embedding (OpenAI today; local models like Ollama planned), and storage so you can skip wiring Pinecone, embedding pipelines, and chunking logic for each project. Key features include a hybrid retrieval score that combines semantic similarity, recency, and user-specified importance (configurable weights), Prisma-based type-safe DB access, Docker Compose for one-command deploys, and admin endpoints for pruning and session management. The standout is a retrieval visualizer dashboard that exposes why a particular memory was recalled, addressing the common “invisible vectors” debugging problem and improving observability for RAG systems. Technical implications: runs on standard Postgres with the vector extension (no vendor lock-in), exposes REST endpoints to store/retrieve/clear memories, and ships with tests and deployment notes for Railway/Vercel. Trade-offs to note: embeddings currently require OpenAI keys unless you add local model support, and scaling/performance will follow Postgres/pgvector constraints. For teams building agent memory or on-prem RAG, MemVault offers a pragmatic, auditable alternative to managed vector DBs.