nornicdb-hermes-memory-gpu
timothyswt/nornicdb-amd64-cuda-bge:latest
https://hub.docker.com/u/timothyswt
bridge
false
https://github.com/orneryd/NornicDB/issues
https://github.com/orneryd/NornicDB
NornicDB is a Neo4j-compatible graph + vector + temporal database and the best long-term memory store for the Hermes agent. This GPU build (NVIDIA CUDA or Vulkan) accelerates embeddings and hybrid search so the Hermes agent gets high-throughput sub-millisecond HNSW vector search, native graph traversal, managed embeddings, reranking + LLM, and policy-based memory decay in one engine. Through Bolt/Cypher and Qdrant-compatible gRPC plus the built-in Heimdall AI assistant, the Hermes agent can persist, auto-link, and recall memory as a self-organizing knowledge graph with MVCC temporal/as-of reads — purpose-built Graph-RAG memory for the Hermes agent at scale. Connection guide (Hermes / Python / Node.js): https://github.com/RapalS/UNRAID_DOCKER_TEMPLATES/blob/main/docs/guides/nornicdb-connection.md
## NornicDB (GPU)
Graph, vector, and temporal database — Bolt/Cypher compatible with Neo4j drivers. **GPU acceleration** for embeddings and inference.
### Image variants
Pick a **version/branch** when adding the container:
| Branch | When to use |
|--------|-------------|
| **cuda-bge** | NVIDIA GPU + BGE-M3 (~4.5 GB). Requires NVIDIA Driver plugin |
| **cuda-heimdall** | NVIDIA + Heimdall AI (~5 GB). Set `NORNICDB_HEIMDALL_ENABLED=true` |
| **vulkan-bge** | AMD/Intel/NVIDIA via Vulkan. Set Extra Parameters to `--device=/dev/dri` (see Prerequisites) |
### Prerequisites
- **CUDA branches (default):** Install Unraid **NVIDIA Driver** plugin. Template adds `--runtime=nvidia`.
- **Vulkan branch:** GPU with `/dev/dri` on the host. The template default **Extra Parameters** is `--runtime=nvidia` (for CUDA). When you select **vulkan-bge**, edit **Extra Parameters** to remove `--runtime=nvidia` and add `--device=/dev/dri` (Unraid does not always carry per-branch parameters across a version switch).
### After install
1. WebUI: port **7474**
2. Bolt: `bolt://YOUR_UNRAID_IP:7687`
3. Health: `http://YOUR_UNRAID_IP:7474/health`
4. **Connection guide (Hermes / Python / Node.js):** https://github.com/RapalS/UNRAID_DOCKER_TEMPLATES/blob/main/docs/guides/nornicdb-connection.md
**Embeddings:** Do **not** bind-mount an empty folder to `/app/models` — that hides the bundled `bge-m3` model.
### Heimdall AI assistant
On **cuda-heimdall** branch: set `NORNICDB_HEIMDALL_ENABLED=true` (bundled GGUF model). Or enable on any BGE branch and configure provider:
- **local** — GGUF in `/app/models` (default)
- **ollama** — `NORNICDB_HEIMDALL_PROVIDER=ollama`, set API URL to your Ollama host
- **openai** — `NORNICDB_HEIMDALL_PROVIDER=openai`, set API key
Open admin UI → helmet icon (Bifrost chat). Guide: https://github.com/orneryd/NornicDB/blob/main/docs/user-guides/heimdall-ai-assistant.md
For CPU-only AMD64 hosts, use **nornicdb-hermes-memory-cpu**. For Apple Silicon, use **nornicdb-hermes-memory-apple-silicon**.
Tools:Utilities Network:Management
http://[IP]:[PORT:7474]
https://github.com/RapalS/UNRAID_DOCKER_TEMPLATES/blob/main/docs/guides/nornicdb-connection.md
Docker Hub pull. CUDA variants need NVIDIA Driver plugin. Recommend 4 GB+ RAM and GPU.
https://raw.githubusercontent.com/RapalS/UNRAID_DOCKER_TEMPLATES/main/templates/nornicdb-hermes-memory-gpu.xml
https://raw.githubusercontent.com/orneryd/NornicDB/main/macos/Assets/NornicDB.iconset/icon_1024x1024.png
--runtime=nvidia --restart unless-stopped
cuda-bge
NVIDIA CUDA + BGE-M3 (~4.5 GB). Default for NVIDIA GPUs.
timothyswt/nornicdb-amd64-cuda-bge:latest
--runtime=nvidia --restart unless-stopped
cuda-heimdall
CUDA + BGE + Heimdall AI (~5 GB).
timothyswt/nornicdb-amd64-cuda-bge-heimdall:latest
--runtime=nvidia --restart unless-stopped
vulkan-bge
Vulkan GPU (AMD/Intel/NVIDIA). Set Extra Parameters to --device=/dev/dri (remove --runtime=nvidia).
timothyswt/nornicdb-amd64-vulkan-bge:latest
--device=/dev/dri --restart unless-stopped
7474
7687
6334
/mnt/user/appdata/nornicdb/data
/mnt/user/appdata/nornicdb/export
/data
7474
7687
true
false
local
qwen3-0.6b-instruct
1024
0.5
-1
local
false