Neo4j Agent Memory

Name: Neo4j Agent Memory
Author: Neo4j Labs

Neo4j Labs

Three memory tiers: short-term (per-session conversation history with vector and text search), long-term (a POLE+O knowledge graph of People, Objects, Locations, Events, and Observations with entity resolution and deduplication), and reasoning (traces of agent tool use and decisions, retrievable for similar-task lookup). Multi-stage entity extraction via spaCy, GLiNER, or an LLM, plus relationship extraction with GLiREL and optional background enrichment from Wikipedia or Diffbot. Two independently versioned SDKs share one memory model and talk either directly to a self-hosted Neo4j over Bolt or to Neo4j's hosted memory service over REST, with cross-SDK behaviour enforced by a companion conformance suite.

Storage: A Neo4j graph database (vector indexes require 5.20 or later), run yourself via Desktop, Docker, or Aura — or proxied through the hosted Neo4j Agent Memory Service, which is still Neo4j underneath but Neo4j-managed.
Retrieval: Hybrid vector plus text search over short-term history; graph traversal and entity-resolution queries over the long-term knowledge graph, including a read-only Cypher accessor that works across both the Bolt and hosted backends; and similarity search over stored reasoning traces.
Self-host: Self-host: moderate
License: Apache-2.0
Pricing: The library and the self-hosted Bolt path are free and Apache-2.0 — you supply your own Neo4j instance, and Community Edition is free. The hosted service requires signup for an API key and its docs reference rate-limit tiers implying paid plans, but no service-specific price list is published; the only pricing link resolves to Neo4j's general corporate pricing page. · Free + paid
GitHub stars: 384
Last release: 2026-05-21
Last commit: 2026-07-22
First catalogued: 2026-07-25

Strengths

A real queryable knowledge graph as long-term memory, with an explicit POLE+O entity model, rather than vector recall with graph branding
Three cleanly separated tiers including reasoning-trace memory, which few catalogued frameworks offer at all
Eight first-party framework adapters, each individually verified, so one memory backend serves several agent stacks
A genuine open self-host path over Bolt against any Neo4j instance — not gated behind the hosted service — which additionally unlocks write-Cypher, geospatial, and air-gapped operation
Cross-language SDK behaviour is enforced by a dedicated conformance test suite, unusual rigour for a Labs project

Watch out

Explicitly labelled Experimental and community-supported in its own README, and not officially backed by Neo4j — this is a Labs project, not a product with support commitments
Hosted-service pricing is undocumented; the only pricing link dead-ends at Neo4j's generic corporate page, which is a budget risk for anyone adopting the zero-infrastructure path
High open-issue count relative to project size, consistent with the Experimental label
The network MCP transport is still flagged as SSE in the CLI, the superseded transport, with no evidence of Streamable HTTP on that surface yet
An unrelated unscoped npm package shares the name, so installing without the @neo4j-labs scope gets the wrong package

Best for

Teams that want an actual queryable knowledge graph rather than vector recall as agent long-term memory
Multi-framework agent stacks needing one shared memory backend across LangChain, CrewAI, ADK, and others
Organizations already running Neo4j who want to reuse that operational expertise

How it integrates

MCP server (local-stdio, remote-http)SDK SDK REST API CLI Framework adapter (langchain, pydantic-ai, google-adk, aws-strands, crewai, llamaindex, openai-agents, microsoft-agent)

Benchmark results

No sourced results yet.

Sources

Last verified 2026-07-25 · updated by discover-frameworks