September 2026 Summaries

SurrealDB University has launched “SurrealDB for AI Engineers,” a free, hands-on eleven-lesson course designed to teach developers how to build the retrieval, memory, and context layers used in modern AI applications. The curriculum covers AI and context engineering fundamentals, vector and full-text search, RAG knowledge bases, hybrid search with reranking, agent memory, text-to-SurrealQL, secure multi-agent context-layer architecture, document chunking, retrieval evaluation, and multi-hop Graph RAG. Built for practical learning, the course requires no prior SurrealDB experience, cloud account, or embedding API, and has students run SurrealDB locally through schemas, datasets, and queries that expose retrieval mechanics. Its progression reflects the development of AI systems from basic semantic RAG toward applications requiring keyword search, structured filters, relationships, persistent memory, permissions, and access to varied data sources, using SurrealDB’s combined structured, document, graph, vector, and full-text capabilities.

SurrealDB Cloud has expanded from AWS-only deployment to Microsoft Azure, initially offering regions in US East 2, Germany, and Brazil, with cloud provider selectable during instance creation. The Azure availability is intended to simplify adoption for organizations standardized on Azure by reducing cross-cloud networking, procurement, latency, and data-egress concerns, while Azure Private Link enables private connectivity within customer virtual networks. Regional deployment supports data-residency needs, including keeping German and Brazilian customer data within those countries, and provides the same security certifications, encryption, service tiers, APIs, backups, versions, and operational behavior available on AWS. Users can begin with a free Azure instance, while the Scale tier offers multi-node, multi-availability-zone production clusters, and an Azure Marketplace listing is planned for future procurement integration.

A SurrealDB 3.x tutorial demonstrates how moving multi-hop graph traversal from application-side breadth-first-search loops into the database can reduce network round trips, transferred data, and latency while enabling filtering during each hop. Using a synthetic GraphRAG dataset of 2,000 entities, 16,000 typed edges, and 800 documents, it shows that unpruned three-hop walks can rapidly expand to hundreds of reachable nodes, making cardinality management more important than individual-hop performance. The recommended approach uses recursive SurrealQL traversal with per-edge filters for relation types and confidence, per-node filters for tenant and visibility constraints, depth ranges, and de-duplicated collection of reachable entities; path output can also provide explanations for why results were retrieved. The tutorial discusses safeguards for dense hubs, noting that per-hop edge slicing bounds expansion but may discard valid routes and can return no results when too few edges exist. It further combines graph traversal with reverse document-mention hops for ranked document retrieval and HNSW vector search for hybrid retrieval, while explaining that normal indexes benefit graph entry points but arrow-based traversal uses direct graph-edge scans rather than secondary edge-table indexes. Benchmark results cited show equivalent 14-entity outputs with server-side traversal requiring one round trip and 241 bytes versus six round trips and about 4,905 bytes for application-side BFS, and the tutorial concludes that pruning at every hop, measuring frontier size, and limiting traversal depth are central practices for scalable graph retrieval.

SurrealDB has introduced the @surrealdb/mastra-ai package for Mastra, a TypeScript AI agent framework, combining a SurrealDB storage adapter with an optional managed Agent Memory provider. The adapter consolidates conversations, workflow snapshots, observability data, embeddings, and native HNSW vector search into a single ACID database, reducing the need for separate databases, synchronization jobs, and multi-system queries. Agent Memory complements transcript storage by extracting durable facts, enabling semantic recall, and building user profiles across sessions while retaining in-process message history and gracefully handling service outages. The two components can be used independently or together, with the storage adapter available for self-hosted SurrealDB or SurrealDB Cloud and the memory provider using the hosted Agent Memory service. The integration requires SurrealDB v3, Bun 1+ or Node.js 22+, and Mastra Core 1.31.0 or later.

Approximate HNSW vector indexes search globally rather than scanning all rows, so filters applied after the nearest-neighbor search, including row-level permissions, subqueries, views, or client-side filtering, can silently remove every returned result when the caller’s documents represent a small share of the index. Putting a visible predicate in the same WHERE clause as the KNN operator produces correct results by pushing the filter into the graph traversal, but can become extremely slow for highly selective tenant or user filters because the index must explore many nonmatching vectors. Increasing K and EF is not a reliable solution, since sufficiently sparse tenants may require traversing nearly the entire index. For small scoped datasets, an exact cosine-distance scan restricted by an indexed ownership or tenant field is both correct and fast, while a hybrid strategy can first use the HNSW index and fall back to that exact scoped scan only when too few results are returned. For multi-tenant systems not yet deployed, isolating each tenant’s data in a separate database or namespace eliminates post-filtering from the vector index, though nonpartitioning filters such as status or language still require query-level handling.

SurrealDB Cloud has launched a Frankfurt, Germany region, providing a second European deployment option alongside Ireland and bringing managed database instances closer to users across the DACH region and continental Europe. The region enables German and EU customer data to be stored and processed in Germany for data-residency needs while retaining the platform’s SOC 2 Type 2, ISO 27001, GDPR, encryption, and network-isolation protections. SurrealDB emphasizes that its combined vector, graph, full-text, and document-query capabilities can reduce the need for separate database and search services, particularly for AI agent workloads. Frankfurt joins regions in the United States, Ireland, Mumbai, and São Paulo, and users can select Europe (Germany) when creating an instance without changing tiers, versions, backups, or existing automation; the Scale tier offers multi-node, multi-availability-zone clusters for high availability.

SurrealDB Cloud is now available in São Paulo, Brazil, allowing customers in Brazil and across Latin America to deploy managed database instances closer to their users. The new South America region is intended to reduce query latency and support Brazilian data-residency requirements by keeping customer data stored and processed within the country, while retaining the platform’s SOC 2 Type 2, ISO 27001, GDPR, encryption, and network-isolation protections. São Paulo joins existing regions in Virginia, Ohio, Oregon, Ireland, and Mumbai, and can be selected during instance creation without changes to existing tiers, versions, backups, or automation beyond specifying the new region. The announcement also highlights SurrealDB’s ability to combine vector search, graph traversal, full-text search, and document reads in one database query, and notes that Scale-tier deployments support multi-node, multi-availability-zone high availability.