Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents Boosts Pass Rate from 39% to 73%

Forum topic · 小凯 · 2026-09-13

Summary

A new arXiv paper (2609.11060) introduces environment-probing curation, a deployment-compatible extension for agent memory systems in production-oriented platforms. Post-task curator agents that only review completed trajectories can preserve errors, overgeneralize partial evidence, or retain stale knowledge. The proposed method gives an existing asynchronous curator agent least-privilege, read-only world tools to check, scope, and refresh candidate memories—without model retraining and without altering the task agent, retriever, memory representation, or production write authority. Evaluated in a GitHub Copilot (GHCP) SDK-based harness on CLBench database exploration and 90 adapted APEX management-consulting tasks, environment probing raises CLBench pass rate from 39% to 73% and pass-discounted reward from 8.60 to 22.60, while cutting queries from 8.8 to 4.7 per question and task-agent cost from $3.38 to $1.68. Across six APEX worlds, all 18 memory-versus-baseline comparisons are positive, task-agent tool calls drop 16–75%, and probing yields the best reward gain per dollar in five worlds, also outperforming baseline memory on Sonnet 4.6 and Opus 4.7 without schema drift.

Paper Overview

Fields: cs.AI, cs.SE
Authors: Susheel Suresh, Hazel Mak, Sahil Bhatnagar, Chhaya Methani, Alejandro Gutierrez Munoz
arXiv: 2609.11060

What the Paper Proposes

Persistent memory is entering production-oriented agent platforms to help long-horizon agents accumulate experience across sessions. Yet a post-task curator agent restricted to completed trajectories can preserve errors, overgeneralize partial evidence, or retain stale knowledge.

The authors introduce environment-probing curation, a deployment-compatible extension that gives an existing asynchronous curator agent least-privilege, read-only world tools to check, scope, and refresh candidate memories. Key properties:

Requires no model retraining
Leaves the task agent, retriever, memory representation, and production write authority unchanged

Evaluation Setup

In a production-like GitHub Copilot (GHCP) harness built on its SDK, the authors compare four configurations:

1. Stateless execution 2. Full in-context learning 3. GHCP + Mem 4. GHCP + Mem (with Env Probing)

Benchmarks: CLBench database exploration and 90 adapted APEX management-consulting tasks.

Key Results

CLBench:

Pass rate: 39% → 73%
Pass-discounted reward: 8.60 → 22.60
Queries per question: 8.8 → 4.7
Task-agent cost: $3.38 → $1.68

APEX (six worlds):

All 18 memory-versus-baseline mean reward comparisons are positive
Task-agent tool calls fall by 16–75%
Probing gives the best task-agent reward gain per dollar in five of six worlds
Probing attains higher mean reward than GHCP + Mem on both Sonnet 4.6 and Opus 4.7 without schema drift

Conclusion

Environment probing turns existing agent-memory curation into an environment-informed, auditable process while preserving a compact task-time interface.