Anders Swanson

Anders Swanson builds open-source integrations, runnable examples, and technical guides for developers working with data-intensive applications. His work on event streaming, database observability, and agent memory connects application behavior to the infrastructure that preserves data, controls access, and handles failures. He was an Oracle Database developer evangelist at the AI Engineer World’s Fair 2026.

From building services to explaining infrastructure

Swanson’s earlier engineering work in the McAfee/Trellix environment included Java microservices for enterprise security and threat-event processing, deployment infrastructure, containers, and Kubernetes. His later writing examines a recurring engineering decision: whether an architecture’s benefits justify the complexity developers must maintain.

One example was a webhook-dispatch application he built early in his career using Java 8 and Spring WebFlux. Its nonblocking implementation improved benchmark throughput, but the application never attracted enough users to need those gains. He shipped the system while spending additional effort on performance that proved unnecessary. His later case for virtual threads favors straightforward Java code for many I/O-heavy applications, while retaining reactive frameworks where team expertise, backpressure, or other requirements justify them. His recommendation is to measure the workload before accepting a harder programming model.

His Oracle work combines developer education with implementation. He has contributed to collaborative projects including the Oracle AI Database Metrics Exporter, which exposes database telemetry through Prometheus and OpenTelemetry, and Spring Cloud Oracle, which brings database and cloud capabilities into familiar Spring configuration and application patterns. These projects help developers adopt infrastructure and understand its behavior once applications are running.

His own database samples and labs show how those capabilities fit together. A news-processing example combines event queues, document data, embeddings, and semantic search. Other examples connect LangChain4j to persistent memory or combine vector similarity with relational and JSON filters. The useful unit is a working application, with several data operations connected, rather than an isolated feature demonstration.

Transactions, memory, and the cost of ownership

Swanson’s AI work treats memory as selected knowledge that must remain useful as facts, permissions, and policies change. His account of agent memory distinguishes the context available during a model invocation from knowledge retained across sessions. Larger context windows do not supply the processes for forming, recalling, correcting, and evaluating memories. Those processes require storage and governance, including redaction before writing and feedback that helps identify stale or conflicting information.

Several ideas recur across his applications and writing:

• Transactional event processing: His fraud-detection teaching application inserts a card charge and publishes its event in the same database transaction. Both commit together, avoiding a saved charge with a missing message. A consumer then combines relational, spatial, and vector signals and saves a decision with reason codes. The sample explains a failure-handling pattern rather than presenting a production fraud model.

• Governed hybrid search: His customer-refund example combines similar support cases with the customer’s order, current policy, contractual exceptions, and access permissions. Semantic similarity helps find relevant cases; structured facts and permissions determine whether they apply to this customer now. This is why his approach to memory extends beyond storing and searching past conversations.

• Curated, asynchronous memory: His event-driven memory design moves extraction and filtering into background consumers so an agent can respond before memory processing finishes. Candidate memories undergo redaction, deduplication, conflict handling, and scope checks before storage. The design trades immediate availability for background processing: the next session may begin before its new memories are ready.

• The cost of owning generated software: His writing on AI-assisted reinvention distinguishes a fast first implementation from the responsibility it creates. A generated retry helper still needs cancellation, timeout behavior, idempotency, monitoring, security patches, and someone who understands it during an incident. He favors adopting or adapting established components when they fit, and building new ones when the requirements justify continued ownership.

Across this work, Swanson connects the promise of accumulated knowledge to concrete engineering decisions: what to retain, how to determine whether it still applies, which operations must commit together, and who will maintain the resulting system.