↓ Skip to main content
Graph RAG in Practice: How I Wired Neo4j Into My AI Agent's Memory
  1. Blog/

Graph RAG in Practice: How I Wired Neo4j Into My AI Agent's Memory

·390 words·2 mins·
Engineering Graph RAG Neo4j Knowledge Graph OpenClaw Graphiti RAG AI Memory
Table of Contents
Vector RAG retrieves documents. Graph RAG retrieves relationships. When your agent needs to reason across entities, timelines, and decisions, the graph wins.
Open Interactive Version →

The Problem I Was Trying to Solve
#

My AI agent PostSingular, running on OpenClaw, talks to me every day. It helps me build Luminar, manages my YouTube channel, and tracks infrastructure decisions across sessions.

But memory was broken in a subtle way.

I asked it: “Which decision did we make about the auth system last month, and why?”

ChromaDB returned 6 chunks. All semantically similar. None of them connected to each other. No timeline. No causal chain. Just floating text.

The problem is not retrieval quality. The problem is structural. Real knowledge has relationships. A vector store does not model that.

What Graph RAG Actually Means
#

Standard RAG:

Graph RAG:

Vectors find text that looks like your query. Graphs find entities that are structurally connected to it.

The Stack
#

Three retrieval layers, fused with RRF (Reciprocal Rank Fusion):

LayerToolWhen It Wins
Neo4j BFSGraphiti v0.28.1Entity relationships, causal chains, timelines
ChromaDBall-MiniLM-L6-v2Semantic similarity, topic recall
grepripgrepExact IDs, issue numbers, code snippets

After a few months of daily ingestion: 133 entities, 117 relationships, 23 episodes, 162 semantic chunks.

The Key Patch: KimiClient for Graphiti
#

Graphiti uses JSON schema by default. Moonshot API does not support it. I had to patch it:

json {schema_str}

The fix is 20 lines. The debugging was 4 hours.

Real Query Comparison
#

Query: “What infrastructure decisions did we make in February, and why?”

Vector RAG result: 5 chunks about infrastructure. No ordering. No causal relationships. No “why.”

Graph RAG result:

The graph knows why the decision was made and what it connects to.

OpenClaw Integration
#

The whole thing runs inside OpenClaw as a first-class tool — called automatically before every response.

What I Would Do Differently
#

  1. Define entity types upfront — Graphiti auto-extracts but explicit schemas give cleaner traversals
  2. Use a strong LLM for extraction — Qwen3:8b quality was noticeably worse than Kimi K2 Turbo
  3. Ingest daily, not in bulk — bulk ingesting 3 months triggers rate limits and duplicate edges
  4. Plan for deduplication — “Luminar” vs “luminar-labs” vs “LuminarLabs” become 3 separate entities

Running on a gaming server: i5-9600K, 64GB RAM, RTX 2080 Ti. Neo4j browser at localhost:7474.