KAIST's PREPING framework builds agent memory without tasks, cuts costs 3x

A persistent challenge in AI agents is the cold-start gap: when deployed in a new environment, they lack task-specific experience to build useful memory. Traditional solutions rely on offline curated demonstrations or expensive online post-deployment interactions. Now, researchers at KAIST introduce PREPING (Proposer-guided Memory Construction), a framework that enables agents to construct procedural memory entirely from self-generated synthetic practice, before ever seeing a real task. The system uses three components: a Proposer that generates synthetic tasks conditioned on a structured control state (proposer memory), a Solver that executes them, and a Validator that determines which trajectories are stored and provides feedback to guide future proposals. This controlled loop avoids the pitfalls of naive synthetic interaction—redundancy, infeasibility, and uninformative trajectories—while selectively updating memory.

Experiments across three benchmarks—AppWorld, BFCL v3, and MCP-Universe—show PREPING substantially outperforms a no-memory baseline and achieves performance competitive with strong playbook-based methods built from offline or online experience. The deployment cost is $2.99\times$ lower on AppWorld and $2.23\times$ lower on BFCL v3 compared to online memory construction. Crucially, the benefit doesn't come from synthetic volume alone but from proposer-side control over feasibility, redundancy, and coverage, paired with selective memory updates. This approach could radically reduce the cost and friction of deploying AI agents in new domains, enabling true zero-shot adaptation at scale.