Neo4j Agent Memory MCP Server

The Neo4j Agent Memory MCP server is a lightweight bridge that lets AI assistants treat a Neo4j graph database as a persistent, semantic memory store. It exposes a set of declarative tools—create, read, update, and connect memories—that allow agents to record facts, link them with meaningful relationships, and retrieve context in natural‑language queries. By delegating all the heavy lifting to a LLM, the server keeps its API surface minimal and predictable: each tool performs exactly what its name describes, while the model decides how to interpret user intent and resolve ambiguities.

For developers building conversational agents that need long‑term recall, this server solves the problem of memory persistence and context richness. Traditional in‑process memory solutions fade after a session ends, but Neo4j persists data across restarts and scales horizontally. The graph structure captures semantic relationships (e.g., KNOWS, WORKS_AT), enabling agents to surface related facts automatically. The search tool’s word‑tokenization strategy returns broader matches, giving the LLM room to rank relevance—an approach that improves as model capabilities grow.

Key features include:

• Persistent, labeled memories: Store any entity type (person, place, project) with arbitrary properties and automatic timestamps.
• Semantic connections: Create typed relationships that encode real‑world associations, enriched with properties such as roles or dates.
• Intelligent search: Natural‑language queries traverse the graph, filter by type or date, and control depth to surface related memories.
• Graph exploration: Agents can navigate relationships, discover new connections, and build richer context for downstream tasks.
• Enterprise‑ready: Support for multiple Neo4j databases and secure authentication via environment variables.

Typical use cases span knowledge‑base assistants, project management bots, and customer support agents that must remember user preferences or past interactions. In a workflow, an agent might store a new client as a memory, link it to previous interactions, and later retrieve all projects the client is involved in by querying the graph. Because each tool is atomic, developers can compose complex behaviors in the model itself without extending the server’s codebase.

What sets this MCP apart is its LLM‑driven intelligence philosophy. By keeping tool logic simple and transparent, the server allows the model to experiment with different inference strategies—entity extraction, conflict resolution, or relationship inference—without needing code changes. As LLMs evolve, the same server can deliver increasingly sophisticated memory handling simply by updating prompts or training data. This combination of graph persistence, semantic richness, and model flexibility makes Neo4j Agent Memory a powerful addition to any AI‑enabled application.