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

In a workshop for AI Engineer, Nacho Martinez, an AI developer advocate at Oracle, argues that agent builders should spend less attention on the frontier model, which they cannot change, and more on the harness, the scaffolding that supplies the model with memory, tools and context. Martinez's claim is that the same model performs better or worse depending on that configuration, which makes the harness the place where builders have leverage. The workshop applies this to an anti-money-laundering agent for a hypothetical mid-sized bank: it runs around the clock over synthetic transactions, looking for patterns such as a large deposit followed by days of heavy cash withdrawals, and writes up suspicious activity for a human to inspect rather than acting on its own.

Most of Martinez's design choices aim to keep the model's working context small because, they say, models solve problems less accurately as the context window fills, whatever the advertised limit. Martinez divides memory into short-term memory, which is the context window itself, and long-term memory with episodic (past conversations and tool logs), procedural (skills and workflows) and semantic (who the user is and what they prefer) parts. Rather than loading every skill and tool into each request, at 3,000 to 5,000 tokens per skill by Martinez's estimate, their toolbox and skillbox patterns store them as vector embeddings and let the agent loop retrieve the few most relevant, three of each by default in the workshop. Skill distillation turns past successful runs into new skill files in the same store, which Martinez presents as a way for a frozen model to get better without retraining. The workshop's loop is capped at eight iterations and 60 seconds of reasoning per step.

Martinez also takes positions on live debates. On files versus databases, they call git worktrees, the usual way to let several agents work on the same files, a workaround for a consistency problem that database transactions solved 30 years ago, while allowing that files suit some jobs and databases others. For database access they favour pre-written SQL queries the agent can call, because generated SQL often trips over differences between database engines, and for retrieval they recommend hybrid search that combines embeddings with keyword search. Their practical advice in the Q&A was to keep the memory layer simple, starting with just short-term and long-term memory, and to build small custom harnesses for specific problems rather than a general rival to existing coding agents.

This is an architecture walkthrough, not an evaluation. The workshop shows no measure of how accurately the agent flags suspicious transactions, and the claim that harness changes alone move model performance is made without a cited study. Martinez describes themselves as a database person who works at Oracle, which is context for their preference for databases. An earlier AI Engineer talk credited to Oracle's Ignacio Martinez, on agent memory and harness engineering, covers related ground.