This Week in Neo4j: Agent Memory, MCP, Skills, Cypher and more

Neo4j announced that every Aura instance now includes a hosted Model Context Protocol (MCP) server, enabling AI clients like Claude Desktop and Cursor to query graph data with zero setup and OAuth-based access controls. The company also highlighted Semvec, a memory engine providing constant-cost semantic memory for multi-agent systems, and skill distillation in NAMS for portable agent memory packages. These updates aim to streamline AI integration with graph databases.

This Week in Neo4j: Agent Memory, MCP, Skills, Cypher and more Staff Community Manager 5 min read Welcome to This Week in Neo4j, your fix for news from the world of graph databases Every Aura instance now ships with a hosted MCP server built in – so Claude Desktop, Cursor and other AI clients can query your graph with zero setup, just OAuth login and schema/read/read-write tools you control. Also in this edition: Semvec, a memory engine giving multi-agent systems constant-cost semantic memory; skill distillation in NAMS, turning a scoped slice of agent memory into a portable, loadable SKILL.md package; and a debugging story from GraphAcademy’s new blog. The Neo4j Discord just got a Server Tag – and we’re celebrating with a Graph Wizard Plushy Contest. Join us on Discord https://dev.neo4j.com/twin4j-joiner , head to adopt-tag, grab the tag, and you’re entered. We’ll pick winners at random every week through the end of September. And while you’re there, check out our Dorm Room Plushy use case too. Happy Graphing, Alexander Erdl COMING UP Livestream : Architecting Trusted Agents: Turning Knowledge Graphs into Secure Policy Engines https://youtube.com/live/6k096 wfHCY on September 15 Conferences : Find us at Step, San Francisco https://stepsf.com/ on August 26-27, SwampUp, NYC https://swampup.jfrog.com/ on September 1-3, Commit Your Code, Plano https://www.commityourcode.com/ on September 3-4, Agentsnexus, Bengaluru https://agentsnexus.io/ on September 4-5, Agentcon, London https://globalai.community/e/x79ncrl7 on September 8, Devopsdays, Portland https://devopsdays.org/events/2026-portland-or/welcome/ on September 8-10, Percona Live, Amsterdam https://perconalive.com/2026-amsterdam/ on September 9-10, AGNTCon + MCPCon, Tokyo https://events.linuxfoundation.org/agntcon-mcpcon-japan/ on September 10-11, Datacon, Calgary https://www.yycdata.ca/datacon/datacon-2026 on September 11 & Big Data, Paris https://www.bigdataparis.com/ on September 15-16 Meetup : Meet us in Malmö, SE https://luma.com/xrnsdv8l , San Francisco, US https://luma.com/sf-meetup-aug26 & Berlin, DE https://globalai.community/e/0pqyb95l on August 25, Tokyo, JP https://jp-neo4j-usersgroup.connpass.com/event/400616/ on August 26, Tokyo, JP https://luma.com/tokyo-hack-night-08-27-26 on August 27, Pune, IN https://www.meetup.com/graphdb-pune/events/315585209/ on August 29, Bangkok, TH https://luma.com/uztd7tkq on September 1, Bengaluru, IN https://luma.com/ox2tsi1p on September 3 & Tokyo, JP https://luma.com/daytonatokyo on September 12 All Neo4j Events : Webinars and More https://neo4j.com/events/ FEATURED COMMUNITY MEMBER: Himanshu Goel https://www.linkedin.com/in/himanshugoel9725/ Himanshu Goel https://www.linkedin.com/in/himanshugoel9725/ Himanshu is an AI Research Engineer and Data Scientist with six years of experience building and deploying production-grade AI, machine learning, and Generative AI systems across finance and healthcare. Connect with him on LinkedIn https://www.linkedin.com/in/himanshugoel9725/ . He has one of the first confirmed sessions at NODES 2026 https://neo4j.com/nodes “Graph-Powered Temporal Validity: Preventing RAG Systems from Citing Dead Regulations”, where he will demonstrate how Neo4j knowledge graphs solve architectural blind spots by modeling regulatory documents as temporal entities with explicit superseding relationships, effective dates, and sunset clauses. AGENT MEMORY: Constant-cost semantic memory for multi-agent systems https://neo4j.com/blog/developer/constant-cost-semantic-memory-for-multi-agent-systems/ Michael Neuberger shares a deep technical walkthrough of Semvec, a memory engine that gives agents persistent semantic memory at constant per-turn token cost, with Neo4j as the shared store for both domain data and conversational state. Built around a healthcare multi-agent scenario, it covers shared cluster memory, drift detection across specialities, cross-shift handover, and consensus voting – all bridged to the graph via a single INVESTIGATED relationship. MCP: Introducing MCP for Aura: Hosted MCP, built into every Aura instance https://neo4j.com/blog/genai/introducing-mcp-for-aura/ Neo4j is rolling out a hosted MCP server built directly into every Aura instance, so AI clients like Claude Desktop or Cursor can query a graph with zero setup – no local server, no credential juggling, just OAuth login through the Aura Console. Jonathan Giffard and Asad Ali give us the details: It ships with schema, read, and read-write tools, letting you control exactly how much access an AI client gets to your production graph. SKILLS: From Agent Memory to Portable Skills https://neo4j.com/blog/genai/from-agent-memory-to-portable-skills/ William Lyon breaks down skill distillation in NAMS Neo4j Agent Memory Service – the process of turning a scoped slice of an agent’s memory graph into a portable, evidence-backed SKILL.md package that a fresh agent can load and run instead of starting cold. It digs into the AIP Agent Instruction Protocol research behind representing skills as typed graphs rather than prose, plus the governance loop that catches and repairs drift when the underlying memory changes. CYPHER: Writing Cypher Queries That Don’t Lose Your Data https://graphacademy.neo4j.com/blog/list-comprehension-for-lost-data GraphAcademy has just launched a blog and Kemi Elizabeth Ojogbede shared a relatable debugging story – three fake test users, one of whom mysteriously vanishes from query results – that turns into a clear lesson on why chained MATCH patterns silently exclude unconnected nodes. She walks through fixing it with list comprehensions and COUNT {} subqueries instead of OPTIONAL MATCH , with a nice payoff: the data that goes missing is usually the data that needs attention most new users, stalled learners, disconnected records . STARTUPS: Subimage https://www.linkedin.com/posts/alexchantavy neo4jgraphtalk-thecube-knowledgegraphs-activity-7493745145989980160-K07P SubImage’s founder told theCUBE at GraphTalk why his startup exists: attackers think in graphs, defenders think in lists. By mapping cloud, identity, and device data into one graph, SubImage gives security teams — and their AI agents – an attacker’s-eye view of exposure, permissions, and attack paths, turning alert floods into clear priorities. It grew out of Cartography, an open-source tool from co-founder Alex Chantavy. CONTINUOUS LEARNING GraphAcademy : Use Neo4j temporal types to filter, compute, and search date-based data https://graphacademy.neo4j.com/courses/cypher-dates-durations Learn on Your Schedule : Go deeper into graph intelligence on Neo4j’s On-Demand webinar library https://neo4j.com/events/on-demand/? event level=intermediate Workshops : Join our virtual classrooms workshops from Fundamentals to GenAI https://neo4j.com/events/? event type=workshops New Webinar : Scale Karpathy’s LLM Wiki with Neo4j GraphRAG – Americas https://go.neo4j.com/WBR-EDU-260825-Karpathy-LLM Registration.html , Europe, Middle East & Africa https://go.neo4j.com/WBR-EDU-260825-Karpathy-LLM-EMEA Registration.html , Asia Pacific https://go.neo4j.com/WBR-EDU-260825-Karpathy-LLM-APAC Registration.html POST OF THE WEEK: Tom Dörr https://bsky.app/profile/tom-doerr.bsky.social Osintgraph maps Instagram social networks by gathering public profiles, followers, and posts into Neo4j, supporting investigation with an AI agent for natural language queries. — Tom Dörr @tom-doerr.bsky.social 9. August 2026 um 18:01 Please share it if you like it