Oracle Rewires AML Agents With Smaller Memory

Nacho Martinez, AI Developer Advocate and Data Scientist at Oracle, built a memory-aware harness for anti-money laundering live on AI Engineer, and argued the scaffold around the model matters more than the frontier model itself.

The demo was a bank workshop: a medium-sized bank, five suspicious transaction types, one pattern a $1 million deposit followed by $50,000 ATM withdrawals for three days.

Martinez defined a harness as all the scaffolding to implement tool use, memory and perception, with the model as a plugin via the chat completions API. "The harness is actually one of the most interesting topics in AI nowadays because these harnesses actually control everything that the AI agent can do," he said, adding that teams should start from scratch and strip 90% of general-purpose bulk when building for a specific problem.

The limit he kept returning to was context. Martinez said performance drops as the window grows, even as limits have expanded from 256,000 tokens to 1 million tokens due to KV cache, and he claimed "The smaller the context window is the better it will solve a problem." Each skill costs 3,000 to 5,000 tokens, the default retrieval is three skills, and tools plus skills can consume 35,000 to 100,000 tokens, or 80% of the request, he said. His Pi harness, he said, injects about 6,000 tokens at the start of every request versus 30,000 to 40,000 tokens for other harnesses, and he placed the user request third in the instruction hierarchy behind the system prompt.

The field setup was concrete. Attendees accessed the lab via GitHub CodeSpaces and workshopwaitingroom.com, with an agent loop capped at 8 iterations and 60 seconds of reasoning per iteration, configurable. Martinez pointed to a Grok model running on Oracle Cloud Infrastructure, name-checked OpenAI, Anthropic, Cloud Code, CodeX and others as harness options, and flagged a new one-hour course with Andrew Ng and DeepLearning.AI on building adaptive agents. He offered to pay for Codespaces himself and invited attendees to the Oracle booth through Wednesday evening.

You have seen this break in production before. Large retrieved toolsets flood the window, the model forgets the deposit that triggered the case, and the worktree workaround reinvents what databases solved 30 years ago, as Martinez put it. Databricks frames the missing piece as context compaction, where the harness decides what stays active and what gets summarized within and across sessions. OpenAI-style semantic caching can cut frontier model costs by 4 to 5 times, Martinez said, but that economy still depends on keeping the active context small and retrievable. Without that compaction, an AML agent either misses the three-day withdrawal sequence or burns tokens re-deriving it.

What had to be true before this demo was frontier chat APIs, embedding stores from 384 to more than 1,000 dimensions, and hybrid retrieval like RRF to pull the right three skills. What still has to happen is hardened cross-session memory that survives the 8-iteration cap without leaking PII across bank cases. Martinez pitched consultants as the perfect user for this workshop because they can fork a lean, custom harness per client typology instead of shipping one bloated general agent.