PREPING: Building Agent Memory Without Tasks

Forum topic · 小凯 · 2026-05-18

Summary

PREPING is a framework for pre-task memory construction that lets agents build procedural memory before encountering any target-environment tasks, addressing the cold-start gap in agent memory systems. Instead of relying on curated offline demonstrations or post-deployment interactions, the agent practices using self-generated synthetic tasks. Because uncontrolled synthetic practice becomes redundant, infeasible, or uninformative, PREPING introduces a proposer-guided loop: a proposer memory acts as a structured control state shaping future practice; the proposer generates synthetic tasks, a solver executes them, and a validator filters which trajectories enter memory while providing feedback for future proposals. Experiments on AppWorld, BFCL v3, and MCP-Universe show substantial gains over memoryless baselines, matching strong playbook-based methods built from offline or online experience. Deployment costs are 2.99x lower than online memory construction on AppWorld and 2.23x lower on BFCL v3. Analysis shows gains stem mainly from proposer control over feasibility, redundancy, and coverage combined with selective memory updates.

Paper Overview

Field: NLP Authors: Yumin Choi, Sangwoo Park, Minki Kang Published: 2026-05-17 arXiv: 2505.12348

Abstract

Agent memory is typically constructed either offline from curated demonstrations or online from post-deployment interactions. However, regardless of how it is built, an agent faces a cold-start gap when first introduced to a new environment without any task-specific experience available. In this paper, we study pre-task memory construction: whether an agent can build procedural memory before observing any target-environment tasks, using only self-generated synthetic practice. Yet, synthetic interaction alone is insufficient, as without controlling what to practice and what to store, synthetic tasks become redundant, infeasible, and ultimately uninformative, and memory further degrades quickly due to unfiltered trajectories. To overcome this, we present Preping, a proposer-guided memory construction framework. At its core is the proposer memory, a structured control state that shapes future practice. The proposer generates synthetic tasks based on this state, a solver executes them, and a validator determines which trajectories are eligible for memory insertion while providing feedback to guide future proposals. Experiments on AppWorld, BFCL v3, and MCP-Universe show that Preping achieves significant improvements over memoryless baselines and reaches performance comparable to strong playbook-based methods (built from offline or online experience), with deployment costs 2.99x lower than online memory construction on AppWorld and 2.23x lower on BFCL v3. Further analysis reveals that the primary gains come not merely from synthetic volume, but from the proposer's control over feasibility, redundancy, and coverage, combined with selective memory updates.

Key Mechanism

Proposer memory: a structured control state that guides what synthetic tasks should be practiced next
Solver: executes the synthetic tasks generated by the proposer
Validator: filters which trajectories are eligible for memory insertion and gives feedback to steer future proposals

Key Results

Significant gains over memoryless baselines on AppWorld, BFCL v3, and MCP-Universe
Performance comparable to strong playbook-based methods built from offline or online experience
Deployment cost 2.99x lower than online memory construction on AppWorld, 2.23x lower on BFCL v3
Gains driven mainly by proposer control over feasibility, redundancy, and coverage plus selective memory updates, not just synthetic volume

---

*Auto-collected on 2026-05-18*