// the find
orneryd/NornicDB
Nornicdb is a distributed low-latency, Graph+Vector, Temporal MVCC with all sub-ms HNSW search, graph traversal, and writes. Using Neo4j Bolt/Cypher and qdrant's gRPC means you can switch with no changes while adding intelligent features like schemas, managed embeddings, reranking+llm, GPU accel, Auto-TLP, Policy-based Memory Decay, and MCP server.
NornicDB is a Go database that combines Neo4j-compatible Bolt/Cypher graph traversal, HNSW vector search, and Snapshot Isolation MVCC in a single engine — no separate vector sidecar. It's aimed at agent memory, Graph-RAG, and knowledge graph workloads where the alternative is Neo4j + Qdrant + an embeddings service glued together. The feature surface is very wide for a project at this star count.
The protocol compatibility story is the real differentiator: existing Neo4j drivers (Python, Go, Node, .NET) connect unchanged, and Qdrant gRPC clients work too. That's an actual migration path, not a promise. The benchmark methodology is more honest than most — raw result JSON files are committed to the repo, hardware is specified, and the HNSW construction speedup (2.7x via BM25-seeded insertion order) is explained with a reproducible article rather than just a badge. The Ebbinghaus-inspired decay/promotion policy system is genuinely novel: you define memory lifecycle rules in Cypher DDL (`CREATE DECAY PROFILE`) instead of baking them into application code or running a separate TTL service. Docker images are tiered sensibly — the minimal CPU image is ~500MB and the Heimdall+GPU full image is 1.1GB, not a monolith.
The badge claims `go >=1.26` which doesn't exist — a small factual error that signals the documentation isn't being proofread carefully, which matters when you're trusting the Cypher compatibility claims. Docker images ship under a personal account (`timothyswt`) with no org backing; fine for experimentation, a real concern for production adoption and long-term maintenance continuity. 'Full Cypher query language support' is contradicted by the roadmap item for 'Neo4j-compatible end-to-end streaming execution + wrapper driver/ORM' listed as planned — the actual compatibility ceiling is unknown without testing your specific queries. 773 stars for a project claiming 12x-52x Neo4j performance wins and UCLouvain academic validation is low; either the benchmarks don't reproduce cleanly in other people's environments, or trust hasn't built yet — either way, production risk is real.