Environment probing lifts Copilot agent pass rate to 73%

On a database-exploration benchmark, a GitHub Copilot agent with persistent memory went from a 39% pass rate to 73% once its curator was allowed read-only probes of the environment before writing anything to memory.

The result comes from an arXiv paper titled "Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents," which starts from a diagnosis: "a post-task curator agent restricted to completed trajectories can preserve errors, overgeneralize partial evidence, or retain stale knowledge." The fix the authors propose is to hand that curator "least-privilege, read-only world tools to check, scope, and refresh candidate memories," and they stress the extension "requires no model retraining and leaves the task agent, retriever, memory representation, and production write authority unchanged."

On CLBench, pass-discounted reward moved from 8.60 to 22.60, queries per question fell from 8.8 to 4.7, and task-agent cost dropped from $3.38 to $1.68. Across 90 adapted APEX management-consulting tasks spread over six worlds, the paper reports "all 18 memory-versus-baseline mean reward comparisons are positive and task-agent tool calls fall by 16–75%; probing gives the best task-agent reward gain per dollar in five worlds."

The abstract publishes no figure for the curator agent's own probing cost, so the task-agent savings are only one side of the ledger.