Memory-Aware Agent Harnesses for Anti-Money Laundering — Nacho Martinez, Oracle

The model is the part you can't change. The harness around it is where you can make real gains. Nacho Martinez walks through the full agent stack (agent loop, short- and long-term memory, context engineering, the toolbox and skillbox patterns, skill distillation and continual learning) and applies it to an autonomous agent that flags suspicious bank transactions around the clock. 🎟️ Join us at the next one: AIE CODE Summit → https://aie.codes Speaker Nacho Martinez — LinkedIn: https://www.linkedin.com/in/jasperan/ GitHub: https://github.com/jasperan Oracle — https://www.oracle.com Workshop: https://workshopwaitingroom.com Timestamps 0:00 Workshop setup 2:07 Intro: Nacho Martinez 2:57 Agenda 4:57 The use case: an AML agent for a mid-sized bank 5:57 The agent stack 8:31 Agentic systems = automation + autonomy 9:16 What is an AI agent? 11:06 Why the harness matters more than the model 13:21 Files vs. databases 15:56 Agent memory: short-term, episodic, procedural, semantic 19:01 Context summarization vs. compaction 19:46 The agent loop 20:51 Toolbox & skillbox patterns 22:30 Skill distillation 24:15 How a frozen model gets better: continual learning 26:10 Instruction hierarchy 27:05 Workshop UI walkthrough 33:35 Q&A: Episodic memory & conflicts 34:35 Q&A: Skills, caching & context budgets 40:49 Q&A: Keep your memory layer simple 43:29 Q&A: Pre-built SQL vs. dynamic queries 46:14 Hybrid search (semantic + keyword) 47:39 Build small custom harnesses