AgenticGraphRAG

The Simplified GraphRAG Stack: Unified Memory Over Fragmented Databases

"Building a Graph RAG System without a Graph Database" sounds like a contradiction in terms.

Recently, an architecture for implementing Graph RAG without a graph database emerged. Store entities, relations, and passages in 3 Milvus collections linked by IDs. Graph traversal becomes ID lookups. No Cypher. No separate infrastructure.

Now a similar principle appears again, but with MongoDB as the unified memory layer instead of Milvus.

The pattern: consolidate all memory - documents, vectors, and graph links - into a single system. Stop fragmenting knowledge across databases.

Here's how it materializes with MongoDB:

The ingestion layer pulls data from URIs, notes, emails, docs. Normalized into a single schema in MongoDB. Raw documents stored durably.

The memory pipeline processes each document: Clean text + metadata. Graph extraction (entities + relationships). Normalization (merge "Abi" vs "Abi Aryan"). 

Embeddings with Voyage AI. Output: knowledge graph objects with triplets, vectors, and metadata.

The unified layer lives in a single knowledge_graph collection with three index types working in parallel: Text index for keyword recall. 

Vector index for semantic recall. Graph links for multi-hop traversal. All in one place.

Retrieval happens through an MCP server exposing GraphRAG as tools. Natural language query for compact retrieval. Deep search for progressive graph expansion. Ingest for new data.

The agent layer (Claude Code) orchestrates reasoning. It decides when to retrieve, when to write memory, when to expand context via graph traversal (2-3 hops).

Skills bridge the gap: Assistant-memory harnesses semantic retrieval. Assistant-learn pushes insights back to memory.

Two independent architects, two database choices, one architectural insight: simplification beats infrastructure sprawl. 

Unified memory is the flywheel. Everything else—retrieval, traversal, reasoning—becomes faster and cheaper when you stop syncing multiple systems.

The emerging pattern: whether Milvus, MongoDB, or something else, the question is not "which database?" It is "can we consolidate graph, vector, and document memory in one coherent layer?"