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

Highlights

• Retrieval serves as the read path while memory functions as the write path for AI agents.

• Promoting nothing eliminates continuity benefits, whereas promoting everything risks poisoning the memory store.

• Oracle divides agent memory into short-term session states and long-term stores covering episodic, procedural, and semantic memories.

• The Oracle AI Database provides ACID guarantees, multi-rights, and a multi-data type engine to manage agent memory without separate vector stores and document databases.

• Context cards organize topics, summaries, relevant information, and recent messages into compact prompt-ready blocks to prevent context bloat.

Timeline

1. 00:56 - 03:36

   Moving from RAG to Memory Systems

   User expectations have evolved from command-led click-and-type interfaces to intent-based software. While foundational models and infrastructure are widely accessible, the data layer belongs to the user and drives differentiation. Agents consist of static models combined with a harness that owns runtime decisions and safety checks.

   • Retrieval-augmented generation provides a useful read path, but memory systems create unique business value.
   • Intent-based systems require continuity across interactions.
   • The AI stack consists of commoditized models and infrastructure layers, leaving the data layer as the primary competitive advantage.
2. 03:36 - 11:53

   Memory Promotion and Decision Gates

   Retrieval alone does not determine what information survives an interaction. A memory promotion pipeline decides whether to admit, reject, or send a candidate record for human review. Testing scenarios in a sandbox environment demonstrate how semantic preferences load automatically in subsequent sessions while unauthorized blanket commands trigger rejection rules.

   • Memory promotion pipelines classify candidates, compare them with existing records, and check required evidence before admission.
   • A financial services demonstration illustrates how semantic preferences and structural outputs persist across brand-new sessions.
   • Guardrails within the harness evaluate requests to prevent undesirable actions like blanket approvals.
3. 11:53 - 18:56

   Architecting Agent Memory and Context Management

   Building memory pipelines manually requires handling summarization, extraction, and token budget constraints. Oracle introduces an agent memory package that simplifies these operations with a single line of code. The multi-data type engine inside the Oracle AI Database handles vectors, documents, and relational data within one system, ensuring end-user isolation and robust search capabilities.

   • Agent memory categorizes data into short-term session states and long-term episodic, procedural, and semantic stores.
   • Database storage provides ACID guarantees, eliminates file tree issues, and handles multiple concurrent writes.
   • Context cards assemble topics, summaries, and recent messages into compact blocks to reduce developer overhead and manage token budgets.