Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents
Susheel Suresh, Hazel Mak, Sahil Bhatnagar, Chhaya Methani, Alejandro Gutierrez Munoz
cs.AI, cs.SE
2026-09-10
Microsoft gives an async curator read-only world tools to verify memories; on CLBench, GitHub Copilot pass rate rises from 39% to 73% and task-agent cost falls from $3.38 to $1.68.
Production agent stacks now keep memory across sessions. After a task closes, a separate curator writes reusable facts and procedures into a store that later sessions can retrieve. Mem0, Claude Managed Agents, and Copilot Studio all follow that pattern.
The evidence available to that curator is incomplete. A finished trajectory is one partial, often mistaken view of the world. It can store an instance answer as if it were a procedure, freeze an inefficient path, or keep a table name that no longer exists. CLBench already showed that schema drift turns those records into stale beliefs. If verification is deferred to task time, the responding agent spends its tool budget rechecking old memory instead of solving the current request.
Microsoft adds environment-probing curation: least-privilege, read-only world tools on an existing asynchronous curator. The task agent, retriever, memory schema, and production write authority stay fixed. No weights are trained.
After the task closes, the curator follows propose–probe–commit. It drafts a candidate from a distilled trajectory, then issues read-only calls to resolve specific uncertainties before create, update, or delete. Probes can separate an incidental answer from a reusable relation, compare a shorter path, test a claimed join on another slice, check preconditions, inspect states the trajectory never visited, or re-query the live environment when drift is suspected. On CLBench the probes inspect tables, join keys, encodings, and post-migration fields. On adapted APEX they inspect file locations, document relevance, and workbooks. Probes never enter the task trajectory, never consume the task budget, and never mutate the world. If no safe read surface exists, the system falls back to trajectory-only curation.
Four GitHub Copilot SDK systems are compared: no memory, full in-context replay of prior trajectories, GHCP + Mem, and GHCP + Mem with probing.
The main runs share gpt-5.4 across task, distiller, and curator. CLBench's 40-question drift schedule changes the SQLite schema after question 20.
| System | Pass | Total reward | Queries / task | Task-agent cost |
| No memory | 39% | 8.60 | 8.8 | $3.38 |
| Full ICL | 61% | 21.39 | 3.0 | $2.01 |
| Mem | 70% | 20.00 | 5.6 | $1.99 |
| Mem + probing | 73% | 22.60 | 4.7 | $1.68 |
Probing uses 1.69M input tokens, versus 2.13M for Mem and 5.42M for Full ICL. On a 30-question no-drift schedule, probing mean reward is 0.748 on Sonnet 4.6 and 0.721 on Opus 4.7, above trajectory-only memory at 0.673 and 0.696.
Across 90 adapted APEX consulting tasks in six document worlds, all 18 memory-versus-baseline reward comparisons are positive, and task-agent tool calls fall from 30.0–71.6 to 13.9–28.4. Probing has the best reward gain per task-agent dollar in five of six worlds. In the world where baseline discovery is most expensive, calls drop from 71.6 to about 19.
The written records change in kind. Trajectory-only curation stores a warning: do not average prcusd, that yields about 52.96, the grader wants 96.23. The probing curator writes the procedure: join itemsg2 to taxng2 on refid, filter category and prc>0, then compare the average. A stale attrsg3 lemma is rewritten to the current productattributesg3.
This is an increment that existing production stacks can absorb. Keep the model, task agent, retriever, record schema, and async CRUD; give the curator read-only connector or MCP access. Probes inherit platform auth and audit, and they sit off the user-facing path. For enterprise agents that must accumulate environment knowledge across sessions, compact indexed lemmas beat growing full-context replay, and write-time checks beat hoping the next task agent will notice a bad rule.
The gain is evidence quality at write time, not extra capacity at task time.
There is no standalone limitations section. The probing-versus-memory gaps often have overlapping confidence intervals, and the authors treat subgroup differences as a mechanism reading rather than a settled effect. One of six APEX worlds shows essentially no probing gain. The extra no-drift lift is larger on Sonnet (+0.075) than on Opus (+0.025); a stronger model that already writes executable lemmas from the trajectory has less room left. Evaluation covers database exploration and consulting document worlds, not code repositories or environments with write side effects. Distillation and curation tokens are billed separately from task-agent cost, so the async bill still has to be paid in deployment.