neo4j-labs/agent-memory

A graph-native memory system for AI agents and context graphs. Store conversations, build knowledge graphs, and let your agents learn from their own reasoning — all backed by Neo4j.

What it solves

AI agents often struggle with persistent, structured memory. This project provides a graph-native memory system that allows agents to store conversation history, build knowledge graphs of entities and preferences, and record their own reasoning processes for later retrieval.

How it works

It implements a three-layer memory architecture backed by Neo4j:

  • Short-Term Memory: Manages per-session conversation history using vector and text search.
  • Long-Term Memory: Maintains a knowledge graph of entities, facts, and preferences, utilizing entity resolution and deduplication.
  • Reasoning Memory: Records reasoning traces and tool usage to allow agents to retrieve similar past decisions and tasks.

The system supports multi-stage entity extraction (via spaCy, GLiNER2.5, or LLMs) and can be deployed as a hosted service (NAMS) or self-hosted via Neo4j Aura (Bolt). It also provides an MCP server to expose these memory capabilities as tools for AI assistants like Claude or Cursor.

Who it’s for

Developers building AI agents who need their agents to remember users across sessions, maintain a structured knowledge base, or audit the reasoning steps taken to reach a conclusion.

Highlights

  • Three-Layer Architecture: Separate storage for conversations, knowledge graphs, and reasoning traces.
  • Flexible Backend: Supports both a zero-infrastructure hosted service (NAMS) and direct Neo4j Aura connectivity.
  • Broad Integration: Native adapters for OpenAI, Anthropic, Bedrock, and Vertex AI, with LiteLLM support for 100+ providers.
  • Framework Support: Integrations for LangChain, Pydantic AI, CrewAI, LlamaIndex, and others.
  • MCP Server: Turn memory functions into tools for MCP-compatible AI assistants.
  • Advanced Graph Ops: Includes entity deduplication, geospatial queries, and the ability to adopt existing Neo4j graphs.

Related

  • Project
    SuanmoSuanyangTechnology/MemoryBear

    A next-generation AI memory management system that emulates human-like cognitive processes to provide agents with dynamic, evolving long-term memory.

  • Project
    doobidoo/mcp-memory-service

    A persistent, shared memory backend for AI agents that uses a knowledge graph and semantic search to maintain context across sessions.

  • Project
    angelnicolasc/graymatter

    A zero-dependency, persistent memory layer for AI agents that provides a self-building knowledge graph and reduces context token usage by up to 90%.

  • Project
    neo4j-labs/create-context-graph

    Create Context Graph is a CLI wizard that scaffolds a full‑stack AI application backed by a Neo4j knowledge graph. Choose a domain (e.g., healthcare), an LLM agent framework (Strands, PydanticAI, LangGraph, etc.), and a memory backend (hosted NAMS or self‑hosted Neo4j). The tool generates a FastAPI backend, a Next.js + Chakra UI frontend, Neo4j schema, optional demo data, and connectors for SaaS sources. It lets you prototype graph‑augmented agents quickly, with optional production‑grade self‑hosted mode for full Cypher and GDS support.

  • Project
    MemMachine/MemMachine

    An open-source long-term memory layer for AI agents that enables persistent storage and recall of user preferences and conversational context across sessions.