MCP Long-Term Memory — GraphRAG Memory Server¶

What It Is¶

A Model Context Protocol (MCP) server that provides GraphRAG-based long-term memory to AI agents. Uses Neo4j for knowledge graph storage, enabling agents to persist, query, and evolve structured knowledge across sessions.

GitHub: https://github.com/null-create/mcp-long-term-memory (Python, new)

Why GraphRAG: Unlike simple vector stores that retrieve semantically similar chunks, GraphRAG builds a knowledge graph — entities, relationships, and communities — enabling multi-hop reasoning and structured recall that vectors alone cannot provide.

Tools Available¶

Tool	Description
`store_memory`	Persist a fact or observation into the knowledge graph
`query_memory`	Natural language query over the knowledge graph
`get_entity_context`	Retrieve all known facts about a specific entity
`find_relationships`	Discover relationships between entities
`summarize_topic`	Generate a structured summary of what the agent knows about a topic

Quick Start¶

# Prerequisites: Running Neo4j instance
docker run -d --name neo4j-mcp-memory \
  -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/password \
  neo4j:latest

# Clone and install
git clone https://github.com/null-create/mcp-long-term-memory.git
cd mcp-long-term-memory
pip install -r requirements.txt

# Configure
export NEO4J_URI="bolt://localhost:7687"
export NEO4J_USER="neo4j"
export NEO4J_PASSWORD="password"

# Add to Hermes
hermes mcp add memory --command "python" --args "server.py" --workdir "$(pwd)" \
  --env NEO4J_URI NEO4J_USER NEO4J_PASSWORD

Manual Configuration¶

{
  "mcpServers": {
    "memory": {
      "command": "python",
      "args": ["server.py"],
      "workdir": "/path/to/mcp-long-term-memory",
      "env": {
        "NEO4J_URI": "bolt://localhost:7687",
        "NEO4J_USER": "neo4j",
        "NEO4J_PASSWORD": "password"
      }
    }
  }
}

Architecture¶

AI Agent (Hermes/Claude/GPT)
    ↓ MCP Protocol
MCP Long-Term Memory Server
    ↓ Neo4j Driver
Neo4j Graph Database
    ↓
Knowledge Graph
├── Entities (people, companies, concepts, projects)
├── Relationships (works_at, depends_on, mentions)
└── Communities (clustered subgraphs for topic modeling)

Business Use Cases¶

Persistent Agent Assistants: Agents that remember user preferences, past decisions, and project context across sessions
Enterprise Knowledge Base: Auto-building knowledge graphs from agent interactions — every question asked and answer found becomes structured knowledge
Multi-Agent Memory: Shared knowledge graph across a fleet of agents — "what did the research agent learn that the operations agent needs?"
Audit Trail: Every agent decision and its context stored as a queryable graph for compliance

Comparison: Vector vs. GraphRAG Memory¶

Feature	Vector Store (e.g., Mem0)	GraphRAG (this server)
Recall	Semantic similarity	Structured + semantic
Relationships	Implicit (chunk proximity)	Explicit (named edges)
Multi-hop Reasoning	No	Yes (graph traversal)
Entity Disambiguation	Weak	Strong (entity nodes)
Query Complexity	Simple retrieval	Complex graph queries
Setup	Simple (no DB)	Requires Neo4j

Limitations¶

Requires running Neo4j instance (operational overhead)
New server — GraphRAG integration may be incomplete
GraphRAG quality depends on LLM extraction quality (garbage in, garbage out)
Not a drop-in replacement for simpler vector memory solutions