Context Cores: Versioned, Portable Agent Memory That Ships Like Code
The Context Core pattern — versioned, promotable knowledge bundles — sits adjacent to the joelclaw docs memory pipeline and vault-reader, which currently have no version pinning or context portability.
TrustGraph treats agent memory like a software dependency. The core concept is the Context Core — a portable, versioned bundle that packages everything an agent needs to know: ontologies, entity graphs, embeddings, source provenance, and retrieval policies. You build it, test it, version it, promote it through environments, pin it in production, and roll it back when something goes wrong. It’s the same lifecycle as a Docker image or an npm package, applied to context.
Most RAG setups dump documents into a vector database and call it done. TrustGraph goes three layers deep: DocumentRAG for fast semantic lookup, GraphRAG for relationship traversal, and OntologyRAG for precision retrieval against a defined domain schema. The graph layer — backed by Apache Cassandra, Neo4j, or Memgraph — gives you structured relationship queries that pure cosine similarity can’t answer. When the question isn’t “find something similar to this” but “find everything connected to this entity two hops out,” the graph layer earns its keep.
The storage layer is a full multi-model database system: tabular, key-value, document, graph, and vectors all coexist in one deployment. Qdrant handles vectors by default, Cassandra handles the graph, and Garage handles object storage. The whole thing deploys with a single npx @trustgraph/config — it generates either a docker-compose.yaml or a k8s resources.yaml and walks you through the rest. TypeScript clients (@trustgraph/client, @trustgraph/react-state, @trustgraph/react-provider) and MCP integration mean existing agent tooling connects without custom glue.
The Context Core pattern sits directly adjacent to where the joelclaw docs memory pipeline and @joelclaw/vault-reader currently live — those systems handle context injection but have no concept of a versioned context artifact you can promote or pin. The idea that you’d cut a v1.2.0 of your agent’s domain knowledge and deploy it the same way you deploy code is worth watching.
Key Ideas
- Context Cores are versioned bundles — ontology + entity graph + embeddings + provenance + retrieval policies — that you can ship between projects and environments like any other build artifact
- Three RAG modes: DocumentRAG (semantic similarity), GraphRAG (relationship traversal), OntologyRAG (schema-grounded precision)
- Multi-model storage by default: Cassandra for graph, Qdrant for vectors, Garage for objects — swappable for Neo4j, Pinecone, Milvus, MinIO, etc.
- Quickstart is
npx @trustgraph/config— generates Docker Compose or k8s manifests, no hand-rolling infra - MCP server integration means agents can call into TrustGraph as a tool without custom adapters
- TypeScript UI libraries for React frontends alongside the REST, WebSocket, Python, and CLI APIs
- Supports all major LLM APIs (Anthropic, OpenAI, Gemini, Mistral, Cohere, Bedrock) plus local inference via Ollama, vLLM, LM Studio
- 3D GraphViz workbench for exploring installed knowledge bases visually
- Observability via Prometheus + Grafana out of the box
- Python-native (PyPI:
trustgraph), actively maintained and trending on Trendshift