- Claude Web sovereign gateway on node03.
- CC sovereign gateway on node02.
- Qdrant and Trilium running per node rather than as a single shared global dependency.
Case Study 02
Snapshot: March 2026
Sovereign Stack, private memory infrastructure with real deployment shape.
This system turns memory from an application feature into a cluster-level service. Claude Web and CC each get their own sovereign gateway, local knowledge store, reranking, and cross-read search paths instead of sharing one vague context bucket.
Phase One
Already deployed across the cluster.
The value here is not the idea of memory. It is the operating shape: separate gateways, separate stores, streamable MCP transport, and explicit cross-read paths where they are useful.
- BGE-large embeddings and BGE reranking running with CUDA.
- Streamable HTTP transport exposed on the MCP endpoint.
- Cross-read tools so one sovereign surface can selectively search the other.
The system treats memory as infrastructure with topology, transport, and ownership boundaries. That is materially different from plugging a vector store into an app and calling it long-term context.
- Perception bridge and typed emotion state.
- Neo4j knowledge graph and graph-enriched search.
- Redis for ephemeral state and wider federation access.
This is the difference between “memory support” and a real sovereign context layer that other agents, services, and interfaces can build on without losing ownership boundaries.
Infrastructure
Designed for federation, not a toy demo.
- Qdrant
- Trilium
- BGE-large-en-v1.5
- BGE-reranker
- Streamable HTTP
- Claude Web gateway
- CC gateway
- Cross-read search
- Cluster nodes
- ETL pipeline
Need memory to behave like infrastructure instead of a plugin?
That usually means topology, ownership, and retrieval quality all need design work.