GraphNews

Building Agentic Memory on Graphs — Reality Check “Agentic memory on graphs” is quickly becoming a buzzword. But in practice, we are still very early. After winning the NVIDIA Hackathon with our VoiceApp solution (https://lnkd.in/dRuiyDKu) — where we used graphs as agentic memory — I hit a wall. The uncomfortable realization: 👉 Triples are not enough. 👉 Named entities are not enough. 👉Just tracking the flow is also not enough, you have Langgraph state for it! These graphs structure data… but they do not necessarily capture meaning. Even when pushing knowledge graphs further, I kept running into the same limitations: Complex, multi-dimensional relationships (especially in causal systems), Context that evolves across time, actors, and layers, dependencies between relationships, edges depending on edges and nodes having subnodes? In real-world client scenarios, this complexity is the norm—not the exception. And traditional graph models struggle to reflect it. 💡 Accenture slogan is Reinvent. The shift in perspective. Everything started to click when I came across the work of Volodymyr Pavlyshyn on metagraphs. Traditional graphs (and even hypergraphs) still fall short when modeling: relationships between relationships, multi-entity interactions, hierarchical and temporal context. You need metagraphs to make it work. Metagraphs break the rigid structure of classical graphs: Edges can connect sets of nodes to sets of nodes. Edges themselves can be treated as first-class entities. Relationships can have other relationships. Everything is a graph, there is distinction any more between nodes and egdes, only the context matters! Mind blowing death of the rigid Ontology! This is exactly what agentic systems need for real memory. If we want agents to: reason across contexts, handle causality, build long-term, structured memory …then we must go beyond: - triples - entity extraction - static schemas Because real-world knowledge is: - non-linear - nested - multi-layered And our data structures need to reflect that. ⚠️ The current gap There is still no mature, out-of-the-box technology that fully supports this paradigm. But what is exciting: 👉 Practical workarounds are offered in Volodymyr's book: from extending graph DBs (e.g., Neo4j-style abstractions) to modeling metagraph concepts directly in SQL or hybrid systems 📚 Final thought We are at the same point with agentic memory as we were with knowledge graphs for AI a few years ago: 👉 The idea is ahead of the tooling. Metagraphs might be one of the missing pieces to move from structured data → machine-understandable meaning. Highly recommend digging into this space. 🔖 Hashtags #AgenticAI #KnowledgeGraphs #ArtificialIntelligence #GraphData #CausalAI #Metagraph #AIEngineering #LLM #DataModeling #neo4j #nvidia

Background For full disclosure: I have spent the last twenty years inside knowledge graphs and their predecessors. Around 2005 I read Sir Tim Berners-Lee's 2001 Scientific American paper on the Semantic Web — the idea that the web should link not just documents but the things documents describe, so

Sovereign Artificial Intelligence, semantic infrastructure, and the map beneath the model

SHACL 1.2’s First-Class Support for RDF Linked Lists

Why Building a Semantic Layer Without a Validated Use Case Is Dangerous  Many organizations build a semantic layer such as an ontology or knowledge graph simply because it feels strategic or forward thinking. Yet without validating real customer questions first, it becomes one of the most expensive anti patterns in data management.  The Core Danger Solution in Search of a Problem  A semantic layer is an abstraction. It defines entities, relationships, and rules. Without a validated use case you cannot decide which entities matter, at what granularity, or which rules must be enforced.  The outcome is predictable   - Over engineering model everything and slow down progress.   - Under engineering miss essential relationships and fail to answer business questions.   Both create structures that look impressive but fail to deliver business value.  The Myth Hidden Insights Will Appear Automatically  Connecting data alone does not create insight. Without business context, knowledge graphs often surface irrelevant or misleading results.  | Risk | Description | |------|--------------| | False signals | Apparent correlations without causal meaning. | | Insight overload | Thousands of trivial findings with no priority. | | Fragmentation paradox | Poor alignment produces inconsistent subgraphs and new silos. | A Healthy Customer Driven Process  Customer Decision Need → Competency Question → Minimal Ontology → Query → Validation → Iterate  The knowledge graph is never the starting point. The real customer need always comes first.  The Bottom Line  - Do not build semantic layers without validated use cases.   - Validation comes from real customer questions that represent decisions.   - A knowledge graph cannot validate itself; it is a reasoning tool, not a measure of value.   - Start small, deliver proof, then expand. Trust is earned through useful answers, not elegant models.  👉 Follow me for Knowledge Management and Neuro Symbolic AI daily nuggets   👉 Join my group for more insights and community discussions [Join the Group](https://lnkd.in/d9Z8-RQd)  #KnowledgeGraphs #OntologyEngineering #SemanticAI #KnowledgeManagement #NeuroSymbolicAI #AIEngineering #ExplainableAI #DataGovernance

A companion to "The Semantic Medallion", building out some of what the original walkthrough deliberately scoped out.

2) Formal ontologies such as RDF/OWL (Resource Description Framework and Web Ontology Language) trace their origins back to the need for multi-dataset

As enterprise AI moves from dashboards to action, the difference between agreeing on numbers and representing business reality becomes…

Isabelle Augenstein ESWC 13 May 2026

How AI agents persist state across sessions, why memory is different from RAG, and the open-source projects building this layer.

Knowledge graphs automatically constructed from text are increasingly used in real-world applications. However, their inherent noise, fragmentation, and semantic inconsistencies significantly...

As requested, these are the slides of the invited talk I gave yesterday in the event "Ontology, AI, and Cybersecurity" organized by Zequan Huang and Antoine Leblanc at the University of Paris I: Panthéon-Sorbonne. Unfortunately, I could not be there in presence but I had a great time interacting with the audience (great questions!). My presentation covers a number of interesting things that have been developed at Semantics, Cybersecurity, and Services (SCS)/University of Twente over the years. This includes UFO(https://lnkd.in/dWeGR7P5) but also work championed by Ítalo Oliveira, Stefano Nicoletti, Thijs van Ede, Rodrigo F. Calhau and Tiago Prince Sales Relevant papers here include: https://lnkd.in/dvZHu5p5, https://lnkd.in/dBgmCtCk, https://lnkd.in/e47bZvRN, https://lnkd.in/eEm_ajj9, https://lnkd.in/dzKnM7zT, https://lnkd.in/eAdcBJcD, https://lnkd.in/e2axrW7Z Great initiative and thanks for the invitation, folks! Hope to be able to visit the Sorbonne soon! #Ontology #ontologies #AI #cybersecurity #security #risk #riskpropagation #riskmanagement #riskpropagationgraphs

How lance-graph turns Lance tables into a Cypher-queryable property graph with graph traversal, vector search, and multimodal payloads in one system.