This paper investigates the construction of procedural memory in AI agents prior to any specific task, using only self-generated synthetic practice. To overcome the redundancy and ineffectiveness of pure synthetic interaction, the authors propose Preping, a proposer-guided framework that shapes future practice.
This article explains why the Nexus Core AI OS team had to implement persistent memory for their AI agents, moving away from an initial stateless design. They encountered bugs where agents forgot information and repeated tasks, leading them to build a system to address these contextual memory issues.
Agent harnesses are becoming the dominant method for building AI agents and are intrinsically tied to their memory. Using a closed harness, particularly one behind a proprietary API, means yielding control over your agent.
