Total Recall: Agent Memory and Harness Engineering — Ignacio Martinez, Oracle
Ignacio Martinez from Oracle presents how to build effective AI agents by constructing a "harness"—a layer of code around a language model that makes it reliable and repeatable.
He argues that model weights are frozen and free to rent, but everything you actually control lives in the layer surrounding it. He organizes this into seven layers with particular focus on where memory physically lives: files are cheap to append to but lack transactional consistency, while databases (via work trees) solve the problem of parallel agents trampling each other. He introduces the concept of umwelt to describe how an agent's perception is bounded by the institutional knowledge you documented, and presents two key ideas: hysteresis variables for how long the harness should try before giving up, and skill promotion where successful workflows are distilled into better versions.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
