September 2026 Summaries

Decision traces are structured, persistent records that explain how AI agents reached decisions by linking outcomes to reasoning steps, tool calls, policies, conversations, and other relevant context. Stored as reasoning memory within a context graph alongside long-term enterprise knowledge and short-term conversation history, they can support explainability, debugging, compliance audits, consistency, learning, and collaboration among multiple agents. Unlike application logs, which track events, or LLM traces, which detail individual model runs, decision traces focus on why a decision was made and can be retrieved as precedents for future similar tasks. Neo4j Agent Memory provides SDKs, framework integrations, and APIs for explicitly capturing, completing, inspecting, and searching these traces in a graph, enabling teams to review agent behavior and give agents reusable decision history.

Neo4j Labs has introduced the experimental `@neo4j-labs/nams-ai-provider`, a community provider for the Vercel AI SDK that gives AI agents persistent, cross-session memory through the hosted Neo4j Agent Memory Service. The package automatically retrieves relevant memories before a model call and stores new information afterward, using a graph-based model of entities, preferences, decisions, histories, and confidence rather than only vector-searchable chat text. It supports provider, middleware, and tool-based integrations, allowing developers to choose invisible automatic memory, visible model-controlled memory, or a hybrid approach, while hooks can enforce that memory is queried and conversation turns are persisted. Retrieval searches several memory sources in parallel, limits prompt injection to avoid overwhelming user requests, and includes fallback searches for semantically mismatched wording. Optional extraction converts stored content into typed entities that can be queried, corrected, merged, and traced across conversations, though hosted API support for graph relationship edges is not yet available. Designed to run through HTTPS without requiring a Neo4j deployment or vector database setup, the package aims to fail safely when memory services are unavailable, but it remains actively maintained experimental software with no service-level or backward-compatibility guarantees.

A Neo4j-based Neocarta semantic layer is presented as a metadata-only graph that helps Text-to-SQL agents discover relevant tables, columns, and join paths across large, messy data catalogs without repeatedly scanning warehouse schemas. Unlike metric-focused semantic layers that define certified measures over curated data, this approach emphasizes cross-platform schema discovery by ingesting warehouse metadata, descriptions, foreign-key relationships, sample values, and embeddings into a graph while leaving actual data in the warehouse. In benchmarks on a 278-table BigQuery Census catalog and a 264-table Databricks legacy lakehouse, graph-based retrieval reportedly improved reliability for smaller models and reduced token use, latency, and warehouse catalog queries for larger models; compact retrieval on Census reduced tokens by 45–81%, while Databricks tests showed 19–54% token reductions across eight Anthropic models. The approach uses semantic search to locate an anchor table and graph traversal to supply related tables and join details, allowing agents to generate SQL after limited retrieval rather than brute-force exploration. The post argues that such a graph can complement rather than replace existing governed metric layers, potentially expanding into an enterprise knowledge layer containing glossary terms, ownership, governance metadata, usage-derived relationships, and agent memory.

Neo4j describes a no-code, multi-agent integration in which Salesforce Agentforce uses the Model Context Protocol (MCP) to send natural-language business questions to Neo4j Aura Agent, which handles graph-specific reasoning and returns grounded recommendations. Using a Northwind retail example, the Aura Agent can identify a customer, analyze purchase history, determine an unavailable frequently ordered product, find suitable in-stock alternatives in the same category, rank them, and provide supporting evidence through deterministic Cypher Templates. Agentforce remains responsible for employee interaction, intent routing, Salesforce permissions, and potential CRM workflows, while Aura Agent owns graph traversal, domain rules, and retrieval logic. Unlike a prompt template, Aura Agent includes planning, tool selection, graph retrieval, and publishing through secured REST or MCP endpoints. The architecture uses machine-to-machine authentication and read-only Aura operations, while keeping graph schemas, ranking rules, privacy constraints, and Cypher logic on the Neo4j side so Salesforce can continue submitting stable business-level requests.

Using a small Neo4j test graph of three learners with different levels of course activity, including one user with no relationships, the post demonstrates how mandatory Cypher relationship patterns can unintentionally exclude unconnected nodes from results. It explains that `MERGE` patterns should contain only stable identity properties, such as email addresses, while mutable details belong in `SET` to avoid duplicate nodes. A query matching users through enrolments and courses returns only users who satisfy every relationship requirement, causing new or inactive users to disappear despite existing in the database. To preserve all users, the recommended approach is to anchor the query on the primary nodes and retrieve optional related information with list comprehensions, `COUNT {}` patterns, or `COLLECT {}` subqueries, which can yield empty lists or zero counts without removing rows. Although `OPTIONAL MATCH` can also include missing relationships, chaining several optional matches can multiply rows and create unnecessary work, whereas separate list expressions avoid duplication. The broader lesson is that dashboards and reports can overlook precisely the unconnected customers, employees, or learners who may need attention, so queries should be reviewed for implicit requirements before concluding that data is absent.

Neo4j’s Document Intelligence can be adapted from its intended GraphRAG and document-analysis use cases to create an AI-supported homework quiz system from text-based PDF study materials. Using a free Aura instance, users define a simple graph model containing Question and Answer nodes, import a document to generate question-and-answer pairs linked to source chunks, and remove pairs that do not contain exactly one answer. A free Aura Agents approach can then randomly retrieve questions, compare student responses with graph-provided textbook answers and context, and assign correctness and quality scores, although it cannot retain progress or prevent repeated questions. For more control, a small Java application and Cypher queries can track already asked questions, use an external LLM API such as OpenAI to evaluate answers on a 0–3 scale, provide explanations and correct answers, and calculate an overall result. The demonstration uses material on the 1657–1658 Swedish-Danish war and illustrates both the potential of structured graphs for tutoring and practical limitations such as single-user state tracking, API costs, and the need to respect copyright when uploading documents.