Agent memory with environment probing: performance gains without retraining

Researchers have introduced a method that equips an existing curator agent with read-only access to world tools, cutting costs in half while doubling the success rate. The paper, published as a preprint on arXiv and not yet peer-reviewed, addresses a known problem in enterprise agent platforms: persistent memory accumulated at the end of each task tends to preserve errors, over-generalize from partial evidence, or retain knowledge that has gone stale.

The proposed solution, "environment-probing curation," grants the existing curator agent — which runs asynchronously and does not block the task agent — minimal read-only privileges to environment tools. The agent can now verify, scope, and refresh candidate memories against the current state of the world without modifying the task agent, the retriever, the memory representation, or write authority to production. No model retraining is required.

The evaluation was conducted in a production-like GitHub Copilot (GHCP) environment built on its SDK. Four configurations were compared: stateless execution, full in-context learning, GHCP with memory (GHCP + Mem), and GHCP with memory and environment probing (GHCP + Mem w/ Env Probing). Two benchmarks were used: CLBench for database exploration, and 90 adapted APEX management-consulting tasks across six distinct "worlds."

On CLBench, adding environment probing raised the pass rate from 39% to 73% and the pass-discounted reward from 8.60 to 22.60. At the same time, the number of queries per question fell from 8.8 to 4.7, and task-agent cost dropped from $3.38 to $1.68. The figures suggest that probing lets the task agent work with cleaner, more relevant memory, so fewer tool calls are needed to reach a correct answer.

On APEX, all 18 averaged comparisons between memory and baseline showed reward improvement. Task-agent tool calls decreased by 16% to 75% depending on the world. In five of the six worlds, the probing configuration achieved the best reward-per-dollar for the task agent. Additionally, the probing configuration attained a higher average reward than GHCP + Mem on both Sonnet 4.6 and Opus 4.7, without schema drift — meaning the memory structure did not distort over time.

As a preprint, the findings have not undergone formal peer review. The environment is described as "production-like," not actual production, and the code and data are available alongside the paper for replication. Latency metrics for the curator agent itself were not published, and it remains unclear how the method behaves when environment tools return errors or partial information. Even so, the approach demonstrates that existing memory curation can be upgraded to an environment-aware, auditable process without touching the narrow task-time interface.