Staff Community Manager
4 min read
Welcome to This Week in Neo4j, your fix for news from the world of graph databases!
Enterprise AI fails because every agent rebuilds business context from scratch, and the fix is a shared, governed graph substrate called Knowledge Layer.
We take an in-depth look into that, plus a benchmark showing how mid-run agent steering recovers up to 14 points of recall in knowledge graph extraction from pharma documents; a full walkthrough of NAMS, the Neo4j Agent Memory Service, covering its three connected memory types; and a two-hour live coding session on Neo4j Aura for a solid on-ramp into graph.
Share your experiences and influence the future of Neo4j products: Join the Neo4j User Research panel! It’s a chance to connect directly with product development teams, get paid compensation, hear about what we are working on and more!
Happy Graphing,
Alexander Erdl
COMING UP!
Livestream:Query. Explore. Visualize. Meet Neo4j Enterprise Studioon July 28 &Building Automotive Parts Intelligence using Graph Memory for Supply Chain Resilienceon July 29Conferences: Find us atStepSF, San Franciscoon August 26-27,Agentsnexus, Bengaluruon September 4-5,AGNTCon + MCPCon, Tokyoon September 10-11 &Big Data, Parison September 15-16Meetup: Meet us inBerlin, DE&Miami, USon July 28,Berlin, DEon August 25,Tokyo, JPon August 26,Tokyo, JPon August 27 &Pune, INon August 29All Neo4j Events:Webinars and More
**FEATURED COMMUNITY MEMBER: **[Frédéric Valentin](https://www.linkedin.com/in/frederic-valentin-ab313920/)
[Frédéric Valentin](https://www.linkedin.com/in/frederic-valentin-ab313920/)
Frédéric heads Wealth Planning at SGPB Europe, teaches at HEC Paris and Sciences Po and builds his own AI-powered wealth and tax tools in Swift and Python.
Connect with him on LinkedIn. He has one of the first confirmed sessions at NODES 2026 “From Hallucination to Citation: A Neo4j-Backed Legal Agent over MCP”, where he will show you how to model multi-jurisdiction tax law in Neo4j; how to design MCP tool surfaces an agent can orchestrate without getting lost; and how to combine vector search with graph traversal so the agent can both find a relevant rule and walk its dependencies.
KNOWLEDGE LAYER: The Enterprise Knowledge Layer
Jesús Barrasa makes a sharp architectural argument: enterprise AI fails not because of the model or the scaffolding, but because meaning is scattered across agent prompts, MCP servers and retrieval pipelines – ten agents, ten private copies of the business, all drifting. His answer is the Enterprise Knowledge Layer: a shared, governed graph substrate that covers ontology, data and memory, which agents query continuously rather than each rebuilding the business context from scratch´every time.
AGENTS: Improving knowledge graph creation in life sciences through agent steering
Miriam Kümmel and Mathis Lucka from Blue Guardrails benchmark a concrete technique – agent steering – against plain-agent knowledge graph extraction from pharmaceutical SmPC documents stored in Neo4j. Instead of front- instructions, an evaluator intercepts the agent mid-run, pinpoints missed nodes, hallucinated edge attributes and incorrect SNOMED codes and injects a correction prompt; the numbers are specific: up to 14 percentage-point recall improvement and 10+ points on edge-attribute F1, with smaller models benefiting most.
AI MEMORY: A Tour of the Neo4j Agent Memory Service (NAMS)
Will Lyon walks through every tab of NAMS (Neo4j Agent Memory Service), a Neo4j Labs cloud service that gives agents three connected memory types – short-term conversation, long-term entity knowledge graph, and reasoning traces – backed by a managed Aura database with async extraction, POLE+O entity typing, SAME_AS deduplication with a human-review queue and continuous background compression into observations and reflections.
##### CYPHER: [Graph Database Neo4j Masterclass](https://www.youtube.com/watch?v=INB17ozuINY)
New to graph databases? This two-hour live coding session by Sudhanshu Kumar covers everything you need to get started with Neo4j Aura: spinning up a free instance, writing your first Cypher queries to create nodes and relationships, fetching and updating records and finishing with vector storage for RAG applications – a solid on-ramp for developers coming from SQL or MongoDB backgrounds.
##### STARTUPS: [Velasight](https://www.linkedin.com/pulse/what-era-we-actually-real-estate-never-abandons-came-before-jones-tzzye)
Velasight is a graph-native decision intelligence platform for institutional commercial real estate. We use GraphSAGE GNNs trained on property ownership networks to deliver probabilistic site selection and underwriting intelligence across multifamily and data center asset classes. Velasight is currently live in Atlanta, Miami, and Amsterdam.
CONTINUOUS LEARNING
GraphAcademy: Discover yourlearning pathand find exactly what you need to build with Neo4jLearn on Your Schedule: Go deeper into graph intelligence on Neo4j’sOn-Demand webinar libraryWorkshops: Join our virtual classrooms workshopsfrom Fundamentals to GenAI** New Webinar**: The next wave of AI: Insights from AI Engineer World’s Fair –Americas,Europe, Middle East & Africa,Asia Pacific
POST OF THE WEEK: Jose Luis Latorre
Please share it if you like it!