Neo4j Knowledge Graph Memory for AI Agents in 2026

TL;DR — Neo4j Labs shipped neo4j-agent-memory in 2026 — a graph-native memory layer for AI agents with three layers in one knowledge graph: short-term (full conversation chains), long-term (entity + preference graph), and reasoning (record of how the agent solved past problems). It plugs into Microsoft Agent Framework, LangChain, Pydantic AI, Google ADK, Strands, and CrewAI.

The technique

Vector memory is fast for fuzzy recall but blind to relationships. Graph memory excels at multi-hop entity queries — "which patients did Dr. Lee see in February who also saw Dr. Park in March, and which of them have BCBS in-network?" That is a 4-hop join, expensive in vector + SQL, trivial in Cypher.

The neo4j-agent-memory schema models three memory layers as connected sub-graphs in one graph:

• Short-term: (:Session)-[:HAS_MESSAGE]->(:Message) chains
• Long-term: (:Person)-[:WORKS_AT]->(:Org), (:Person)-[:PREFERS]->(:Preference), etc.
• Reasoning: (:Plan)-[:STEP]->(:Action)-[:USED_TOOL]->(:Tool) with outcome metadata

flowchart LR
  C[Conversation turn] --> EX[NER + relation extractor]
  EX --> ST[(Short-term graph)]
  EX --> LT[(Long-term graph)]
  P[Agent plan] --> RM[(Reasoning graph)]
  Q[Query] --> CY[Cypher router]
  CY --> ST
  CY --> LT
  CY --> RM
  CY --> A[Agent]

How it works

On every turn, an entity extractor runs (multi-stage: spaCy/GLiNER for cheap; LLM fallback for hard cases). Entities get upserted with embedding-based dedup ("Sagar S" -> "Sagar Shankaran"). Relations are extracted with a small LLM. New facts are appended; conflicting facts trigger a reconciliation prompt.

Retrieval is a typed Cypher query, not a similarity search. The agent has tools like graph_query_neighbors(entity), graph_find_path(a, b), graph_get_preferences(user). For fuzzy recall the agent calls a separate vector index over node text + descriptions.

The reasoning layer is the underrated piece: every plan + tool-call + outcome is logged. The agent can query "have I solved a problem like this before?" and lift the playbook.

CallSphere implementation

CallSphere uses Neo4j as the cross-entity memory layer for verticals where relationships matter most:

• Healthcare: (:Patient)-[:HAS_PLAN]->(:InsurancePlan)-[:IN_NETWORK_WITH]->(:Provider) and (:Patient)-[:PRESCRIBED]->(:Medication) for allergy/interaction checks.
• OneRoof real estate: (:Buyer)-[:WORKING_WITH]->(:Agent)-[:BROKERAGE]->(:Brokerage), (:Listing)-[:IN]->(:Neighborhood)-[:ZONED_FOR]->(:School).
• UrackIT IT helpdesk: (:Incident)-[:ON]->(:Service)-[:DEPENDS_ON]->(:Service) for blast-radius reasoning.

37 agents · 90+ tools · 115+ DB tables · 6 verticals. $149/$499/$1499, 14-day trial, 22% affiliate. Vertical pages: /industries/it-services, /industries/real-estate.

Build steps with code

from neo4j_agent_memory import AgentMemory
from neo4j import GraphDatabase

driver = GraphDatabase.driver("bolt://localhost:7687", auth=("neo4j", "***"))
mem = AgentMemory(driver)

# Write
mem.add_long_term(
    user_id="patient_4421",
    text="Patient is allergic to penicillin and amoxicillin.",
    extract_entities=True,
)
mem.add_short_term(session_id="s_99", role="user", content="Same kid as last visit")

# Read with Cypher
with driver.session() as s:
    rows = s.run("""
      MATCH (p:Person {id: $uid})-[:ALLERGIC_TO]->(d:Drug)
      RETURN d.name AS drug
    """, uid="patient_4421").data()

1. Run a multi-stage entity extractor: spaCy/GLiNER first, LLM only on miss.
2. Always keep an embedding index alongside for fuzzy recall.
3. Version every fact with asserted_at and source properties.
4. Index hot relationship patterns; Cypher without proper indexes is slow.

Pitfalls

• Entity drift: same person becomes 3 nodes. Hard-enforce dedup with embedding match + alias rules.
• Schema explosion: 200+ relation types makes querying chaotic. Cap at 30–50.
• Cost of LLM extraction: at scale the extractor dominates the bill. Use cheap statistical NER first.
• No conflict policy: when "user lives in NY" and "user lives in Seattle" both exist, the agent picks at random unless you implement temporal reconciliation.

FAQ

Graph or vector? Both. Graph for entity-heavy queries; vector for fuzzy recall.

Neo4j or Memgraph? Neo4j for ecosystem and labs (agent-memory, GenAI integrations); Memgraph for raw query throughput.

Cypher complexity? Mid. A senior engineer is productive in a week.

Cost? Neo4j Aura starts at hobby tier; self-host community edition is free.

See it on /demo? Yes — try a multi-hop query like "find providers in-network for both my plans."

Sources

• Neo4j Agent Memory - Neo4j Labs
• neo4j-labs/agent-memory GitHub
• Building an AI Agent with Memory: MS Agent Framework + Neo4j - Medium
• LLM Memory Graph: Connecting Knowledge - AI Agent Memory
• Meet Lenny's Memory: Building Context Graphs for AI Agents - Neo4j