PREPING: Building Agent Memory Without Tasks via Proposer-Guided Synthetic Practice

Forum topic · 小凯 · 2026-05-18

Summary

PREPING is a research paper (arXiv:2505.12348) addressing the cold-start gap in agent memory: when an agent enters a new environment with no task-specific experience, traditional offline or online memory construction offers no help. The authors study pre-task memory construction, where an agent builds procedural memory using only self-generated synthetic practice before observing any target tasks. Naive synthetic interaction fails because tasks become redundant or infeasible and memory degrades from unfiltered trajectories. PREPING solves this with a proposer-guided framework: a proposer memory acts as a structured control state shaping future practice; the proposer generates synthetic tasks, a solver executes them, and a verifier selects qualifying trajectories for insertion while feeding back guidance. Experiments on AppWorld, BFCL v3, and MCP-Universe show significant gains over memoryless baselines and performance comparable to strong offline/online experience-based methods, with 2.99x lower deployment cost on AppWorld and 2.23x lower on BFCL v3. Analysis shows gains stem from control of feasibility, redundancy, and coverage plus selective memory updates, not merely synthetic volume.

Paper Overview

Field: NLP
Authors: Yumin Choi, Sangwoo Park, Minki Kang
arXiv: 2505.12348

Problem

Agent memory is typically constructed either offline from curated demonstrations or online from post-deployment interactions. Regardless of construction method, an agent faces a cold-start gap when first introduced to a new environment without any task-specific experience.

The paper studies pre-task memory construction: whether an agent can build procedural memory *before* observing any target-environment tasks, using only self-generated synthetic practice. However, synthetic interaction alone is insufficient — without controlling what to practice and what to store, synthetic tasks become redundant, infeasible, and ultimately uninformative, and memory degrades quickly from unfiltered trajectories.

Method: PREPING

PREPING is a proposer-guided memory construction framework with three components:

1. Proposer memory — a structured control state that shapes future practice; the proposer generates synthetic tasks based on this state. 2. Solver — executes the generated synthetic tasks. 3. Verifier — determines which trajectories are eligible for insertion into memory and provides feedback to guide future proposals.

Results

Evaluated on AppWorld, BFCL v3, and MCP-Universe:

Significant improvements over memoryless baselines
Performance comparable to strong script-based methods built from offline or online experience
2.99x lower deployment cost than online memory construction on AppWorld
2.23x lower cost on BFCL v3

Further analysis reveals that the main gains come not merely from synthetic volume, but from the proposer's control of feasibility, redundancy, and coverage, combined with selective memory updates.

Original Abstract (excerpt)

> 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...

Full paper: arxiv.org/abs/2505.12348