August 2026 Summaries

Karpathy’s LLM-maintained wiki concept aims to preserve synthesized knowledge from source documents rather than repeatedly deriving it through raw-file vector RAG, but the approach can suffer from context growth, duplicate content, stale summaries, and inefficient navigation in file-based Markdown collections. The post argues that representing Markdown files, sections, reading order, and links in a Neo4j graph enables scalable progressive disclosure, allowing agents to traverse relationships, inspect backlinks, find paths, rank central documents, and identify communities without scanning an entire corpus or relying solely on similarity search. It presents ki as an open-source, deterministic tool that indexes Markdown into a disposable Neo4j graph without modifying source files or requiring LLM or embedding work during ingestion, offering commands for outlines, full-text search, ordered content retrieval, and direct Cypher queries. Citing research from Newcastle University’s National Innovation Centre for Data, the post says agents combining graph navigation with vector search outperformed vector-only agents on complex factual questions, while also describing a planned ki theme-detection feature based on graph community detection. It concludes that graph-based retrieval could support personal knowledge bases, shared team documentation, agent memory, and structured enterprise knowledge layers by making explicit document structure and links navigable as collections grow.

Neo4j’s enterprise knowledge layer series begins by modeling an organization’s operating structure as a governed graph that makes business units, teams, functions, roles, people, reporting lines, and accountability queryable over time. Using the fictional AcmeBank, the article illustrates how stale org charts and fragmented systems can prevent teams and AI agents from reliably determining who owns a business area, such as SMB lending, while a live graph can provide current and historical answers. The proposed model ingests mapped data from systems of record, applies entity resolution, provenance, schema validation, and selective validity dates, and represents reporting and ownership primarily through stable role nodes rather than individual people. Built in Neo4j using graph types and Cypher, the example dataset demonstrates queries for current role holders, past accountability, reporting chains, and escalation paths. The operating structure is presented as a foundational layer for governed enterprise AI, with later layers intended to connect organizational roles to processes, data products, systems, tools, and agents.

Neo4j’s Going Meta Season 3 examined how ontologies can support the creation, validation, governance, memory, and actions of AI agents. Episodes demonstrated automated ontology development using Neo4j’s data modeling MCP server and agent skills, with competency questions, source data, reusable models, visualizations, OWL/Turtle serialization, and iterative quality checks. The season also challenged the idea that ontologies alone prevent LLM hallucinations, finding that plain-language descriptions can perform as well as formal ontology serializations and emphasizing post-generation schema validation. It introduced ontology-quality measures covering structural, requirement-alignment, pragmatic, and logical criteria, alongside SHACL-derived graph validation reports. Other episodes showed ontologies enabling generic agent tools, shaping short-term, long-term, and reasoning memory, preserving ontology-design decisions, and improving extraction from conversations through configurable NLP, local-model, and LLM pipelines. The final episode explored graph-structured agent skills and Neo4j’s NAMS service, which can distill successful agent interactions into versioned, auditable skills, detect drift, and reduce token use. Season 4 is planned to expand from individual ontology techniques to the broader AI-and-ontologies landscape.

GraphSAGE in Neo4j Aura Graph Analytics addresses the cold-start limitation of transductive embedding methods such as Node2Vec and FastRP by learning an inductive function that can generate embeddings for previously unseen nodes using their features and graph neighborhoods. The post demonstrates this approach with a synthetic e-commerce graph containing customers, products, purchases, and referrals, training a multi-label GraphSAGE model with customer and product properties projected into a shared feature space. The trained model is stored in Aura Graph Analytics’ model catalog, allowing it to be retrieved and applied without retraining when 50 new customers and their initial purchases are added. Embeddings for both existing and new customers are then written to AuraDB and used in a filtered K-nearest-neighbors similarity search that matches new customers with comparable established customers. The example illustrates a workflow in which models are retrained periodically while embeddings for incoming entities can be generated on demand, supporting growing graphs and recommendation-related use cases.

GraphSAGE is an inductive graph embedding method that converts nodes and their relationships into numeric vectors for machine learning applications such as recommendations, classification, and similarity search. Using a movie platform example, it samples a fixed number of a node’s neighbors, aggregates their features, combines that summary with the node’s own attributes, and applies learned weights, biases, and activation functions to produce embeddings. Training can be supervised, where embeddings predict known labels such as whether a movie is a hit, or unsupervised, where connected nodes are encouraged to have similar representations while unconnected nodes are pushed apart. By stacking layers, GraphSAGE incorporates information from increasingly distant neighbors, enabling embeddings to reflect multi-hop relationships such as users with shared viewing histories. Although the examples use single values for clarity, real implementations use high-dimensional vectors and weight matrices to capture many kinds of similarity simultaneously. Its central advantage over older node-embedding approaches is that it learns a reusable aggregation function, allowing meaningful embeddings for newly added nodes such as movies, users, or products without retraining the entire model.

Neo4j’s startup guide for Aura recommends first installing the Neo4j CLI and agent skills so coding agents can access current platform guidance, then using the getting-started skill to provision a database, create a schema, load sample data, and generate an application. It explains that Aura sizing should account for graph size, indexes, query patterns, and concurrent demand, while outlining Free, Professional, and Business Critical tiers and directing users to current documentation for detailed recommendations. The guide highlights GraphRAG, the MCP server, Graph Data Science, and persistent agent memory as tools suited to retrieval, agent integration, analytics, and conversational applications. It also warns startups to avoid duplicate credit applications, understand that credits expire 12 months after code issuance, time renewal applications carefully, and assign multiple Aura organization owners. Additional recommendations include using GraphAcademy courses, following secure driver and ingestion practices, load testing before launch, enabling query logging, using secure connections and access controls, and seeking help through Aura tools, Neo4j technical teams, documentation, community forums, and the support portal.

Neo4j’s August 21, 2026 weekly update highlights new graph-database and agentic AI developments, led by a hosted Model Context Protocol server now included with every Aura instance, allowing clients such as Claude Desktop and Cursor to access graphs through OAuth with configurable schema, read, and read-write permissions. Featured technical content includes Semvec, which uses Neo4j to provide persistent, constant-cost semantic memory for multi-agent systems; Neo4j Agent Memory Service skill distillation, which packages portions of an agent’s memory into portable, evidence-backed SKILL.md files; and a GraphAcademy lesson on avoiding missing Cypher query results caused by chained MATCH patterns. The edition also profiles AI researcher Himanshu Goel and his planned NODES 2026 session on using temporal knowledge graphs to prevent RAG systems from relying on outdated regulations, while covering SubImage’s graph-based cybersecurity platform and an OSINT tool that maps public Instagram networks in Neo4j. Upcoming livestreams, conferences, meetups, training resources, and a Discord plushy contest round out the community news.

Neo4j’s integration with IBM watsonx Orchestrate demonstrates two approaches to building graph-enabled AI agents: a declarative native agent that queries a Neo4j companies knowledge graph through the Model Context Protocol and a code-based LangGraph agent that adds persistent cross-session memory through Neo4j Agent Memory Service. The native agent uses Orchestrate-managed connections to inject Neo4j credentials into a locally executed MCP server, avoiding separately hosted infrastructure, while also supporting curated Python tools for predefined queries such as finding a company’s investors. Its behavior is configured through YAML instructions that guide tool selection, schema inspection, and read-only Cypher queries, although the described ADK release has a command-line limitation for importing native agents with toolkits. The LangGraph implementation supplements graph querying with a memory loop that recalls relevant stored facts, generates responses using graph tools when needed, and persists new user information to NAMS for later extraction and retrieval. This allows the agent to personalize answers across distinct conversations, though newly submitted facts may not be immediately searchable because entity extraction occurs asynchronously.

Neo4j’s guide describes a secure integration between Microsoft Copilot Studio, Okta, and Neo4j that preserves each user’s identity when an AI agent queries graph data. Rather than using a shared database service account, Copilot obtains an Okta OAuth access token for the signed-in user, passes it through a Neo4j MCP server deployed on Google Cloud Run, and Neo4j directly validates the token and maps Okta groups to database roles. This approach prevents an agent’s dynamically generated Cypher queries from exposing data beyond a user’s authorized scope, addressing the “confused deputy” risk associated with shared credentials. The setup requires existing Neo4j Okta SSO, a separate Okta application for Copilot Studio, access-token claims for email and groups, exact role mappings, and Cloud Run deployment configured without database usernames or passwords. The guide also covers optional vector search, with embedding credentials stored in Google Secret Manager, Copilot OAuth configuration and tool synchronization, and a staged verification process for validating tokens, MCP tools, Neo4j access, vector search, and agent responses.

Graph technology models data as nodes, relationships, and properties, allowing organizations to store and traverse the connections among entities such as orders, products, suppliers, customers, and systems. The post argues that this connected-data model supplies enterprise AI agents with memory, relevant context, and multi-hop reasoning that conventional vector retrieval and relational joins may not provide effectively, supporting more grounded and explainable responses through knowledge graphs and GraphRAG. It distinguishes native graph databases, which store relationships directly and can support deep traversals, from graph capabilities layered onto row-based databases, while positioning graphs as a knowledge layer that complements relational databases, lakehouses, document stores, and vector search systems. Citing a 2026 IDC study, Neo4j says knowledge-graph grounding reduced hallucinations by an average of 44% in studied deployments, and it highlights applications in fraud detection, supply-chain analysis, recommendations, customer data integration, compliance, network operations, and identity management. The post forecasts growth in the graph database market and promotes Neo4j’s platform, AuraDB cloud service, Virtual Graph capabilities, and GraphAcademy training as ways to adopt the technology.

Neo4j’s Meta Knowledge Graph (MKG) is presented as a harness-agnostic shared context and memory layer for AI agents that keeps enterprise data in its governed source systems while storing metadata, pointers, and learned operational knowledge in a graph. Its authored semantic layer, implemented through Neocarta connectors and MCP tools, maps diverse sources into a common database-schema-table-column model enriched with technical, business, and increasingly operational metadata, enabling agents to search by business meaning before querying underlying systems. Its earned memory layer captures session events through lifecycle hooks and processes them into append-only episodic observations, reusable per-tool query-error patterns that can provide fixes during the same failed turn, and selective durable learnings about projects or users. The design emphasizes provenance, tool-specific error handling, typed storage for structured external data, protection against prompt injection by excluding tool outputs from free-text learning extraction, automated deduplication and contradiction checks, and human review for ambiguous project memories and all user-scoped facts before they can affect a versioned agent persona. The post notes that future work will focus on higher-order consolidation, compression, and decay for project memories.

A Neo4j Developer Blog post describes a Microsoft Fabric and Aura Graph Analytics pipeline used to examine whether professional sumo has shifted away from traditional belt-based grappling toward pushing and hybrid styles. Using 78,710 Makuuchi and Juryo bouts from 2000 onward, the workflow ingests API data into a Fabric Lakehouse, classifies 82 winning techniques into six style families, creates wrestler style vectors and prestige-weighted victory relationships, and imports the resulting graph into AuraDB. Rank-weighted PageRank identifies wrestlers who defeated strong opponents, while K-nearest-neighbor similarity and Louvain community detection identify seven style archetypes based on winning methods; results are then written back to OneLake for SQL and BI analysis. The findings indicate that belt-and-throw specialists declined from about 23% of wrestlers in the 2000s to roughly 16% in later decades, while two-way hybrids rose to nearly 18%, and pushing increased in Makuuchi as throwing declined. The analysis also suggests that the shift reaches top-ranked wrestlers, Juryo remains relatively more belt-oriented than Makuuchi, and some wrestlers change styles as they move into the top division.

Neo4j’s Aura Graph Analytics (AGA) is presented as a managed, on-demand environment for applying Graph Data Science algorithms beyond common uses such as fraud detection, helping teams uncover connectivity patterns that tabular analysis may miss. The workflow begins with identifying suitable graph problems and algorithms, potentially using Aura’s AI assistant or other LLMs to interpret schemas, translate tabular data into graph models, and generate code. Users can validate ideas through Bloom’s visual interface, Cypher procedures in Aura’s Query tool, or the GDS Python client for iterative analysis, machine learning workflows, and non-Neo4j data sources. Data can be projected from Neo4j through flexible Cypher or faster native projections, or constructed from Pandas and Spark DataFrames in standalone sessions, with results written back to Neo4j or exported to external systems. The post also emphasizes estimating memory needs, selecting appropriately sized sessions, scheduling Python-based production jobs through tools such as Airflow, using asynchronous steps for long-running workflows, relying on session TTLs for cleanup, and periodically reassessing capacity as graph data grows.

Neo4j’s weekly update highlights new resources and tools for graph-based AI, with recordings from the AI Engineer World’s Fair emphasizing ontologies as formal logical guardrails for more reliable agentic systems. Community contributions include AgentMemory for .NET, which adds Neo4j-backed short-term, long-term, and reasoning memory to Microsoft Agent Framework and Semantic Kernel applications, alongside an evaluation toolkit for measuring memory quality. A practical GraphRAG tutorial combines Spring AI, Neo4j, and Goodreads data to demonstrate vector search and graph traversal while addressing implementation pitfalls, and Neo4j Bloom now supports whole-graph projections for database-wide algorithms such as centrality and community detection. The update also profiles identity threat detection startup Icite, promotes GraphAcademy courses and webinars, recognizes community member Hamza Farooq’s work in NLP and multimodal systems, and lists upcoming livestreams, conferences, and meetups.

Neo4j’s Agent Memory Service (NAMS) is presented as a graph-based system for converting agents’ accumulated conversations, organizational knowledge, reasoning traces, and tool-use histories into portable, executable SKILL.md packages. Its skill-distillation process scopes relevant memory, identifies recurring successful procedures, derives a structured execution graph from recorded traces, and uses an LLM only to synthesize source-cited descriptions and completion criteria, while deterministic checks validate provenance, coverage, coherence, schema compliance, and PII handling. Distilled skills remain linked through graph relationships to the evidence supporting each claim, require human review before publication, and can be consumed through Claude Code, MCP, REST, or other compatible agent frameworks. The system also monitors changes in underlying memory and tool outcomes for drift, enabling targeted repair of affected steps or requiring full re-distillation when evidence is no longer valid. NAMS adopts concepts from the Agent Instruction Protocol, which represents skills as typed, schema-validated graphs rather than prose instructions, an approach whose cited evaluation reported higher task pass rates and faster execution than freeform skill documentation.

A context graph serves as a persistent memory system for AI agents, integrating long-term enterprise knowledge, short-term conversation history, and reasoning memory to ensure consistent and informed decision-making. By preserving and traversing the connections between memories, it enables agents to gather full context, produce accurate answers, and explain their decisions while leaving an audit trail for workflows. Unlike knowledge graphs, which provide semantic structure and business meaning, context graphs connect domain knowledge to conversation states and decision traces, allowing AI agents to access relevant context easily. Neo4j's Agent Memory facilitates the building of context graphs by connecting conversations, facts, tool usage, and reasoning traces into a queryable structure, enhancing agents' ability to maintain and utilize memory across sessions and tasks. This approach supports multi-agent systems by providing a shared memory space, improving explainability, reducing error likelihood, and optimizing token use, ultimately making AI agents more reliable and effective in long-running workflows.

The blog post discusses the implementation and benefits of Semvec, a system that provides constant-cost semantic memory for multi-agent systems, particularly in healthcare settings. It highlights how Semvec and its extension, Semvec Cortex, enable a single agent to maintain persistent semantic memory using minimal Python code, scaling it across multiple agents with the help of Neo4j as the central repository for domain and conversational data. The system improves efficiency by reducing token costs and processing time, making it feasible to deploy in environments with multiple agents. The integration with Neo4j facilitates seamless querying and persistence of conversational states, including crucial features like drift detection, consensus voting, and cross-shift handover. Semvec's architecture ensures that memories are both scalable and reliable, with Neo4j serving as a single source of truth that bridges domain knowledge and conversational data. The blog also provides practical insights on embedding management, drift detection, and the advantages of using Neo4j's capabilities for maintaining a cohesive data environment.

At the "Hack the Video Agent Context Graph" event, hosted by HackerSquad and AWS Builder Loft, 151 participants engaged in a challenge to develop projects using AWS Strands Agents, OpenAI, TwelveLabs, and Neo4j. The event focused on transforming raw video data into a context graph for agent reasoning, emphasizing efficient use of Neo4j without treating it as a mere storage solution. Jeremy Adams provided a starter repository to streamline development and foster innovation. The top three projects showcased unique implementations: Roopa Gangadhar Devihosur's "MealPrep Video Context Graph" utilized Neo4j to extract detailed culinary information from videos, creating a nutrition dashboard and enabling cross-video ingredient comparisons; Alton Alexander's "Rehearsal" converted video recordings into a coaching tool, scoring delivery and linking weaknesses to exemplary clips using a structured graph of speaking techniques; and Rishabh Bansal, Aditya Das, and Jerry Wen's "ZooVision" offered a monitoring solution for zoo enclosures, generating welfare graphs from overnight footage with a focus on evidence provenance and deterministic analysis. These projects illustrated how a shared foundational toolset and starter code could lead to diverse, innovative solutions across different domains.