← All speakers

Wojciech Pluta

Conference affiliation: Director of AI Technical Product Marketing · Oracle · 2026

Wojtek Pluta is a director of technical product marketing at Oracle, where he helps developers build applications with Oracle AI Database. His work on agent memory addresses a specific problem: deciding which parts of a conversation should become instructions or information that an application can use again. A formatting preference, a customer record and an approval policy need different rules, even when they arrive through the same chat interface.

On this page

From interactive cloud demonstrations to stateful applications

In 2022, Wojtek worked with Victor Martin, Bogdan Farca and Stuart Coggins on augmented- and virtual-reality experiences for a Raspberry Pi cluster. Their work formed part of a larger project led by Chris Bensen. The team connected interactive digital representations of the physical cluster to databases, streaming services and Kubernetes infrastructure; an iPad could become a window onto the installation’s operating data.

Wojtek also co-presented developer sessions on eSports analytics and machine-learning data extraction. His more recent writing examines how databases support AI applications that retain information across interactions. The central distinction is between preserving a conversation and deciding what the application should carry forward from it. A log records what someone said; memory requires decisions about what that statement means, whom it concerns and when it should be revised.

Deciding what an agent should remember

In his stateful AI architecture guide, Wojtek distinguishes policies, preferences, facts, completed experiences and execution traces. Each has different requirements for storage, retrieval and revision. A policy requires explicit authoring, a factual claim needs provenance, and a preference can follow a different acceptance rule. Before an observation becomes a reusable record, the application must classify it, establish who may use it and apply the appropriate verification. Each prompt is then assembled from selected records rather than an indefinitely growing transcript.

His mortgage-agent example makes the consequences concrete. An operator asks to see missing documents at the top of a brief. That preference survives a fresh session and applies to another applicant, while the document status still comes from the new applicant’s own file. A blanket instruction to approve applications is rejected. The example separates three things an agent might otherwise conflate: how the operator wants information presented, what is true of an individual case and who can change lending policy.

Personalization, simpler pipelines and shared context

Wojtek’s collaborative tutorials explore different parts of this design problem:

  • Let a writing profile evolve conservatively. In a tutorial with Allen Helton, previous posts provide examples of a writer’s work, while a structured voice profile describes how the assistant should write. Periodic reflection on newer posts updates that profile cautiously, so a few unusual pieces do not redefine the writer’s style. The assistant adapts through stored context without changing model weights.
  • Use a language model where the task needs one. In a document-processing tutorial with Sandro Volpicella, similarity to labeled examples classifies invoices, purchase orders and delivery notes. A language model enters later to extract structured fields. Keeping original files, extracted text, vectors and JSON in one database reduces the services the application must coordinate and retains the material needed to check its output.
  • Make personal context available across clients. A second-brain tutorial with Linda Haviv exposes shared personal context through the Model Context Protocol. Different AI clients can use the same underlying information, with visibility controls and human approval for write tools. This moves accumulated context beyond a single assistant interface while preserving restrictions on who can read or change it.

1 conference talk

Loading more talks and highlights…

References