Oracle open-sourced an AI developer hub that's worth studying seriously.
A technical repository with reference implementations, benchmarked notebooks, and guided workshops, built around one idea: Oracle AI Database as a unified memory core for AI agents.
If you're learning how to build agents beyond the vector DB tutorial, this is the repository to go deep on.
Here's what's inside:
π± Application reference implementations:
πΉ finance-ai-agent-demo - financial AI agent running vector, graph, spatial, and relational queries on a single Oracle AI Database simultaneously
πΉ agentic_rag - multi-agent RAG with Chain of Thought, processing PDFs, websites, and code repos
πΉ FitTracker - fitness platform built live during a webinar using Oracle 26ai JSON Duality Views FastAPI Redis
πΉ tanstack-shoestore - natural language database queries using Oracle 26ai Select AI
πΉ oci-generative-ai-jet-ui - full-stack with Oracle JET, Kubernetes, and Terraform
π Notebooks that go deeper than most tutorials:
πΉ 11 cognitive architectures in one interactive demo - CoT, ToT, ReAct, Self-Reflection, Least-to-Most, Decomposed Prompting
πΉ Vector keyword hybrid search in a single SQL query - not three calls to three systems
πΉ 6 types of persistent memory implemented and compared
πΉ Filesystem vs database agent memory - compared directly
πΉ F1 Miami 2026 GP strategy using real FastF1 data - SQL, hybrid search, JSON documents, and property graph all in one Oracle 26ai database
π§ Agent Memory package (OAMP):
πΉ One converged engine replacing vector DB key-value store relational store
πΉ Benchmarked against naive flat-history memory - 80 turns, 3 agent variants, token cost latency quality measured
πΉ Three core primitives: users/agents, memories, threads
πΉ End-to-end examples for OpenAI Agents SDK, Claude Agent SDK, and LangGraph
πΉ Deep Research Agent - genome exploration with Tavily OAMP for durable findings across sessions
πΉ Mortgage Approval in LangGraph - OAMP persists applicant data so failed runs resume where they stopped
π« 3 workshops, each self-contained with Codespaces environment and Oracle AI Database pre-configured:
πΉ Information Retrieval to RAG - 5 retrieval strategies over 200 ArXiv papers
πΉ From RAG to Agents - retrieval as tools, persistent session memory
πΉ Agent Memory - 6 memory types, context engineering, with vs without memory compared
Multicloud samples: AWS, Azure, Google Cloud, MongoDB API.
Presented at DevWeek SF 2026 and the DeepLearning[.]AI AI Developer Conference, April 2026.
Open source.