NornicDB: Graph + Vector + Temporal Database
Main Purpose
NornicDB is a distributed, low-latency database combining graph, vector, and temporal capabilities in a single system. Implements Neo4j-compatible Bolt/Cypher protocols while adding vector search, historical reads via MVCC, and AI-native features.
Core Value Proposition
Targets workloads requiring simultaneous graph traversal, vector retrieval, and historical truth -- particularly AI agent memory, knowledge graphs, and Graph-RAG systems. Rather than bolting vector functionality onto a graph database, NornicDB unifies these as co-equal execution paths.
Key Features
Protocol & Compatibility: - Neo4j Bolt protocol and Cypher query language support - REST, GraphQL, and gRPC (Qdrant-compatible) interfaces - Drop-in replacement for existing Neo4j applications
Intelligent Capabilities: - Native vector search with GPU acceleration (Metal/CUDA/Vulkan) - Memory decay systems using Ebbinghaus-Roynard four-layer decomposition - Auto-relationship generation via embedding similarity, co-access patterns, and temporal proximity - 950+ APOC functions
Data Management: - Snapshot Isolation with MVCC for repeatable reads - Canonical graph ledger model for temporal validity and audit trails - Historical reads and as-of queries - Conflict detection at commit time
Advanced Features: - Heimdall AI assistant for natural language database interaction - Knowledge-layer scoring with profile-driven decay and promotion - Retention policies with GDPR erasure tracking - Plugin system for custom extensions
Performance
LDBC Social Network Benchmark (M3 Max, 64GB): - Message content lookup: 6,389 ops/sec vs Neo4j's 518 (12x faster) - Recent messages (friends): 2,769 ops/sec vs 108 (25x faster) - Avg friends per city: 4,713 ops/sec vs 91 (52x faster)
Hybrid Retrieval (vector + 1-hop graph traversal): ~859 microseconds P50 (HTTP)
Deployment
Multiple Docker image variants from minimal (~148 MB) to full with embeddings + AI assistant (~1.1 GB ARM64).
Technology
Go (90.4%), MIT licensed with defensive patent non-assertion grant.