Ignacio Martinez presents an agent-harness workshop built around a simulated bank and synthetic transactions. The focus is the surrounding system: memory, tools, context assembly and the loop that coordinates them. The exercise is an educational example, not evidence of production anti-money-laundering accuracy or regulatory compliance.
Ignacio Martinez distinguishes short-term context from longer-term episodic, procedural and semantic memory. He discusses when files or databases may fit, including consistency problems with concurrent writes. These are design tradeoffs rather than a claim that one storage approach is always correct.
Ignacio Martinez describes retrieving relevant tools and skills dynamically instead of placing every definition into each request. He also discusses distilling useful procedures from prior work and combining keyword with semantic retrieval. These mechanisms update the harness and its stored information; they do not necessarily change the underlying model's weights.
Ignacio Martinez shows a bounded demonstration loop and a dashboard of suspicious findings for later human inspection. In questions, he emphasizes starting with a simple memory design and adapting retrieval to the task. His broad performance claims about context size, caching and search are explanations in a workshop, not universal guarantees validated across every model or workload.
Watch on YouTube
