Preping: Building Agent Memory without Tasks

A new framework called Preping addresses the cold-start gap in agent memory by constructing procedural memory before any task-specific experience. Agent memory is typically built offline from curated demonstrations or online from post-deployment interactions, but agents face a cold-start gap when introduced to a new environment without task-specific experience. Preping uses self-generated synthetic practice to build memory, but synthetic interaction alone is insufficient due to redundancy, infeasibility, and uninformative tasks, as well as memory degradation from unfiltered trajectories. The framework employs proposer memory, a structured control state that shapes future practice, and a Proposer generates synthetic tasks conditioned on this memory. The paper is available on arXiv under identifier 2605.13880.