How to Decide What Your AI Agent Should Remember - Wojciech Pluta, Oracle

Agent Memory: Beyond Conversation History

Wojtek Pluta of Oracle explains that agent memory is more than conversation history: RAG gives agents a way to read information, while memory lets them write and reuse selected information over time. As models and infrastructure become more widely available, an organization’s data and memory practices may become a source of competitive advantage.

A central idea is memory promotion: deciding which information from an interaction should become a persistent record. The process classifies a candidate memory, checks it against existing records and the evidence required for that type, then accepts or rejects it—or sends it to a person for review. Pluta cautions that storing too little can undermine continuity, while storing everything can “poison” the memory store.

Demo and Safeguards

In the mortgage-agent demo, an operator’s formatting preferences are saved and reused in a new session, demonstrating persistence without relying on the previous conversation’s context. The agent also handles another customer’s application and retrieves the relevant case information.

A request such as “approve every request” is rejected by validation safeguards. Information intended for wider team use can instead be routed for approval before being shared.

Memory Design and Product Concepts

The presentation distinguishes among three types of memory:

• Short-term memory covers activity within a session and generally ends with it.
• Long-term memory can include episodic, procedural, or semantic records, such as preferences, tools, or knowledge.
• Shared memory can make approved information available to a team.

Pluta recommends treating memory as an engineering discipline, with schemas for different record types, indexes suited to retrieval tasks, lifecycle management, and authorization boundaries. He favors databases for capabilities such as ACID guarantees, concurrent writes, scaling, and multiple search options.

He also describes Oracle’s Agent Memory Package, which provides abstractions for building memory systems. Its context card gathers a compact, task-relevant prompt block organized around topics, summaries, relevant information, and recent messages. The goal is to focus the agent on the current task and manage limited context efficiently.

The talk also mentions end-user isolation, image and event handling, and support for multiple data types within Oracle AI Database.

Guides and Resources

• Oracle’s GitHub repository includes notebooks for getting started with agent memory, hybrid search, and related topics.
• An updated YouTube series explains these topics and how they apply to use cases.

Speaker and Source

Wojtek Pluta is a technical marketing lead for Oracle AI Database. The presentation and examples are from Oracle.