PREPING cuts agent deployment costs 3× with pre-task synthetic memory

PREPING, a memory construction framework from researchers at KAIST and Yonsei University, lets AI agents build procedural memory before encountering real tasks. The system generates synthetic practice scenarios, executes them, and selectively stores useful trajectories—sidestepping the cold-start problem that typically forces agents to learn from scratch in new environments.

Most agent memory systems rely on curated demonstrations collected offline or on interactions gathered after deployment. PREPING instead uses a three-component loop: a Proposer generates synthetic tasks conditioned on a structured control state called proposer memory, a Solver attempts those tasks, and a Validator decides which trajectories merit storage while feeding back guidance to shape future proposals. The control state prevents the system from practicing redundant or infeasible scenarios, and the validator blocks low-quality trajectories from polluting memory.

On benchmarks and cost

The team tested PREPING on AppWorld (a mobile app interaction benchmark), BFCL v3 (function-calling tasks), and MCP-Universe (a multi-domain agent environment). On AppWorld, PREPING matched the performance of playbook-based methods built from offline or online experience while cutting deployment cost by 2.99× compared to online memory construction. On BFCL v3, the cost reduction was 2.23×. The paper reports that synthetic volume alone doesn't drive the gains—proposer-side control over feasibility, redundancy, and coverage, combined with selective memory updates, accounts for most of the improvement over a no-memory baseline.

The preprint (2605.13880) is available on arXiv. The authors note that unfiltered synthetic interaction degrades memory quality quickly, and that the proposer memory state is the key mechanism for steering practice toward informative scenarios.