Descriptions:
Eli Schilling, a cloud architect at Oracle, led a hands-on workshop at AI Dev 26 in San Francisco on building structured memory systems for autonomous AI agents. The session covered five distinct memory types — conversational history, knowledge base, workflow toolbox, entity summary, and tool log — each backed by a dedicated Oracle Database table, with corresponding LangChain tools and a Tavily integration for live research queries.
The core demo built a research paper assistant capable of ingesting large academic corpora and managing context intelligently across extended agent sessions. Schilling introduced a two-harness model: an inner ReAct loop (Reason, Act, Observe) that runs until a stop condition is reached, and an outer flywheel that wires together a memory manager, context assembly, sub-agent coordination, and post-session memory extraction. A key result shown is that memory-managed agents maintain stable, bounded token usage as sessions grow longer, while naive stateless agents show unbounded token consumption — a practical cost and latency argument for structured memory over raw context stuffing.
All workshop code is available in Oracle’s AI Dev GitHub repository, including a Docker container configuration for running Oracle Database locally and Jupyter notebooks covering the complete implementation. A companion free course on DeepLearning.AI goes deeper on the same material. Schilling also previewed a follow-up 4-hour coding deep dive in the Bay Area focused on advanced agent memory workloads for practitioners who want to extend the Oracle AI Database and LangChain stack further.
📺 Source: DeepLearningAI · Published May 20, 2026
🏷️ Format: Hands On Build
