PaperlayerView exact revision on arXiv

Source-linked AI summary

Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents

Susheel Suresh, Hazel Mak, Sahil Bhatnagar, Chhaya Methani, Alejandro Gutierrez Munoz

arXiv:2609.11060v1cs.AIcs.SE

TL;DR

Trajectory-only memory curation can preserve incorrect, partial, or stale lessons because completed tasks provide limited evidence about the environment. The paper adds least-privilege, read-only environment probing to an asynchronous curator without retraining or changing the task-time memory stack. Across CLBench and APEX, probing improves reward and reduces repeated task-agent interaction and cost, while retaining a compact production interface.

  • Problem

    Trajectory-only curation can preserve errors, overgeneralize partial evidence, or retain stale knowledge because a trajectory cannot establish that a lesson is correct, general, or current.

  • Method

    Environment-probing curation gives a post-task asynchronous curator least-privilege, read-only world tools to check, scope, reenact, and refresh candidate memories before writing.

  • Results

    Across CLBench and adapted APEX, environment probing improves reward while reducing repeated environment interaction and task-agent cost, with the advantage persisting across evaluated model families.

  • Takeaways & Limitations

    Environment probing turns agent-memory curation into an environment-informed, auditable process while preserving the existing task agent, retriever, record schema, and production CRUD lifecycle.

  • Takeaways & Limitations

    Representative trajectory-only records still show incorrect aggregations, incomplete relations, and stale field names, motivating probing as a response to this evidence boundary.

Abstract

from arXiv · show

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. We 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. It requires no model retraining and leaves the task agent, retriever, memory representation, and production write authority unchanged. In a production-like GitHub Copilot (GHCP) harness built on its SDK, we compare stateless execution, full in-context learning, GHCP + Mem, and GHCP + Mem (w/ Env Probing) on CLBench database exploration and 90 adapted APEX management-consulting tasks. On CLBench, probing raises pass rate from 39% to 73% and pass-discounted reward from 8.60 to 22.60 while reducing 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 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. Probing also attains higher mean reward than GHCP + Mem on both Sonnet 4.6 and Opus 4.7 without schema drift. Environment probing therefore turns existing agent-memory curation into an environment-informed, auditable process while preserving a compact task-time interface.

1 Introduction

Persistent memory can help enterprise agents accumulate experience, but trajectory-only curation may preserve incorrect, partial, unverifiable, or stale lessons. Environment-probing curation adds read-only checks by an asynchronous curator while preserving existing task-time components and authority boundaries.

  • 1 Introduction: Trajectory-only evidence can memorize answers, inherit inefficient paths, assert unverifiable scope, miss unvisited states, or become stale.Deferring verification to task time makes the responding agent spend tool calls checking uncertain memories.
  • 1 Introduction: Environment probing lets a post-task curator check claims, test scope, reenact procedures, and refresh stale entries before writing memory.The curator receives the completed trajectory and grade after non-writing distillation, then probes with read-only world tools.
  • 1 Introduction: The work connects environment-informed curation to constructive episodic memory and interactive learning, where interaction can provide competence unavailable from passive observation.The cited accounts motivate checking and recombining experience beyond retrospective records alone.
  • 1 Introduction: The deployment-compatible design requires no retraining and leaves the task agent, retriever, memory representation, and production write authority unchanged.Only the asynchronous curator gains least-privilege, read-only access to existing connectors or MCP tools; probes remain off the user-facing critical path.
  • 1 Introduction: 39% to 73% pass rate and 8.60 to 22.60 reward on CLBench show the reported benefit of environment probing.The introduction also reports lower queries and task-agent cost, while the broader evaluation covers CLBench and adapted APEX tasks.

2 Related Work

Prior agent-memory systems store, retrieve, and refine factual or procedural experience, but their post-task curation remains bounded by recorded evidence. This paper adds read-only environment checks as an orthogonal way to validate and scope memories before writing.

  • 2 Related Work: Agent-memory research transfers persistent facts about users or environments and procedures learned from prior execution while keeping model weights fixed.The paper organizes prompt-based systems by these two transfer targets.
  • 2 Related Work: Factual-memory systems implement create–store–recall lifecycles through episodic retrieval, memory tiers, summarization, reconciliation, and multi-term storage.Examples include Generative Agents, MemGPT, MemoryBank, Mem0, and MemoryOS.
  • 2 Related Work: Procedural-memory systems distill reusable behavior through verbal lessons, successful exemplars, success–failure contrasts, replay buffers, templates, or executable skills.The cited systems vary in the granularity and form of the reusable behavior they retain.
  • 2 Related Work: Post-task curation still mainly operates over records, trajectories, grades, and usage signals, so it cannot recover unobserved states or verify rules after environment drift.The limitation applies despite advances in memory content, representation, retrieval, and rewriting.
  • 2 Related Work: Environment probing is presented as orthogonal: factual or procedural curators can check claims, reenact procedures, inspect omitted states, test scope, and refresh stale knowledge before writing.The added capability changes the evidence available to curation rather than the memory content or representation alone.

3 Method

The method separates task execution, trajectory distillation, and asynchronous memory curation. Its environment-probing extension gives the curator targeted, read-only tools to validate, revise, narrow, delete, or skip candidate records before commitment.

  • 3.2 Post-Task Memory Curation: The task agent is a fresh session with environment tools and read-only retrieved memory, while the curator acts only after task closure and receives the trajectory, feedback, and relevant records.The curator is separate from the task-agent session, cannot see future tasks, and does not answer the completed task.
  • 3.2 Post-Task Memory Curation: The curator is the only agent allowed to mutate persistent memory through read, create, update, and delete operations.Committed records become available when the next task begins.
  • 3.2 Post-Task Memory Curation: A non-writing distiller converts the raw trajectory into a compact view containing the task, retrieved memories, decisive observations, procedures, unresolved assumptions, and answer.The raw trajectory remains available as evidence, while the distiller has no feedback, memory tools, environment tools, or write authority.
  • 3.2 Post-Task Memory Curation: The curator is instructed to maintain small, reliable, transferable, actionable records and may store scoped facts, procedures, relations, conventions, and warnings rather than incidental answers.It evaluates evidential support, transfer value, scope, actionability, and overlap before deciding how to update memory.
  • 3.3 Environment-Probing Curation: Representative trajectory-only records preserve incorrect aggregations, broad maps without needed relations, or removed field names, leaving future agents to establish actionability and currency.These examples make the retrospective evidence boundary concrete in CLBench.
  • 3.3 Environment-Probing Curation: Environment probing keeps the task agent, retriever, distillation, curator, schema, and CRUD policy unchanged while adding only a safe read-only environment-tool subset after closure.In CLBench and APEX, probes inspect environment artifacts to evaluate proposed memories rather than solve future tasks.
  • 3.3 Environment-Probing Curation: The curator follows a propose–probe–commit process that investigates uncertainties before creating, revising, narrowing, deleting, or skipping a memory.Probes can test reusable relations, shorter procedures, preconditions, omitted states, and suspected environment drift.

4 Experiment Setup

The experiments compare four GHCP memory conditions on CLBench and APEX using repeated seeded runs and run-level confidence intervals. Evaluation measures strict task success, tool use, tokens, reward, and cost under fixed task-agent accounting.

  • 4 Experiment Setup: Four systems are compared: GHCP (No Memory), GHCP + Full ICL, GHCP + Mem, and GHCP + Mem (w/ Env Probing).Full ICL prepends prior trajectories, while both memory systems expose memory_read and probing adds curator-only read-only tools.
  • 4 Experiment Setup: CLBench uses a primary 40-question drift schedule in which a SQLite schema changes after question 20.The benchmark models database exploration under evolving schema conditions.
  • 4 Experiment Setup: Primary CLBench and APEX experiments use gpt-5.4, with five paired seeded runs per CLBench configuration and five stateful or three stateless APEX runs.Results are reported as run-level means with 95% Student-t confidence intervals.
  • 4 Experiment Setup: Pass-discounted reward uses binary task pass scores and benchmark budgets of B = 15 SQL-query calls for CLBench and B = 100 Archipelago tool calls for APEX.A failed task receives zero, while a passing task receives more reward when it uses fewer task-agent tool calls.

5 Results

Across CLBench and adapted APEX, memory improves reward and reduces task-agent interaction, while environment probing further improves efficiency and reward by validating and refreshing records before reuse.

  • 5.1 CLBench: Pass rate rises from 39% to 61–73%, while total reward rises from 8.60 to 20.00–22.60 and task-agent cost falls from $3.38 to $1.68–$2.01.Queries also fall from 8.8 to 3.0–5.6 per question.
  • 5.2 Adapted APEX: Across six APEX worlds, all 18 memory-versus-baseline reward gains are positive, and probing achieves the best task-agent reward gain per dollar in five worlds.The largest tool-call reduction occurs in world 941eba66, from 71.6 to 17.7–19.3 calls.
  • 5.1 CLBench: Indexed memory preserves reusable schemas, relations, file maps, and procedures with 2.13M input tokens, compared with 5.42M for full in-context learning.Probing reduces this further to 1.69M input tokens while reaching the highest CLBench pass rate and reward.
  • 5.1 CLBench: Probing raises drift-CLBench reward from 20.00 to 22.60, lowers queries from 5.6 to 4.7, and reduces task-agent cost from $1.99 to $1.68.On no-drift CLBench, mean reward rises from 0.673 to 0.748 on Sonnet and from 0.696 to 0.721 on Opus.
  • 5.3 Why Memory and Probing Work: Probing curates executable records specifying source tables, join keys, filters, grain, and current schema rather than only failed answers or search locations.Matched cases connect validated relations and workbook maps to fewer queries or tool calls.
  • 5.3 Why Memory and Probing Work: Probing improves over trajectory-only memory in five of six APEX worlds, but its gain is −0.04 in world 941eba66 and subgroup differences remain unresolved because uncertainty intervals overlap.The authors interpret probing as most useful when joins, workbook locations, or procedures remain unresolved.

6 Conclusion

Environment probing addresses the inability of trajectory-only memory to establish whether lessons are correct, general, or current by giving an asynchronous curator read-only world access before memory writes.

  • 6 Conclusion: Read-only probes improve reward while reducing repeated environment interaction and task-agent cost across CLBench and adapted APEX.The improvement persists across GPT-5.4, Sonnet 4.6, and Opus 4.7.
  • 6 Conclusion: The deployment boundary leaves the model, task agent, retriever, record schema, and asynchronous CRUD lifecycle unchanged; only the curator gains least-privilege read-only connector or MCP access.Probes remain off the critical path, add no write authority, and inherit platform authentication and auditing.

A Implementation Details and Extensions

The implementation uses fresh SDK-managed sessions for task agents, distillation, and curation, while an external index persists memory across tasks and probing adds only read-only curator tools.

  • A.1 Runtime Architecture and Session Lifecycle: Each task starts a fresh Copilot CLI session configured through the Python SDK with a model, instructions, tools, MCP servers, permissions, and callbacks.Callbacks expose ordered messages, tool calls, and observations for the harness.
  • A.1 Runtime Architecture and Session Lifecycle: After task completion, a distiller condenses the raw trajectory, then a fresh curator receives the distilled trajectory, feedback, and nearby memories before the next task begins.The external index persists across tasks, while the task agent has no memory-write tool.
  • A.1 Runtime Architecture and Session Lifecycle: The trajectory-only and probing curators share the model, inputs, memory CRUD tools, and core prompt, but probing additionally supplies two instructions and the task’s read-only tool subset.CLBench uses database queries; adapted APEX uses read-only MCP tools.
  • A. Implementation Details and Extensions: The runtime mirrors production-oriented Copilot Studio through configured instructions, models, memory, callable tools, skills, connected agents, connectors, and MCP resources.This decomposition supports enterprise APIs and file-like API or document resources.
  • A. Implementation Details and Extensions: Probing reuses registered connectors or MCP tools with least privilege, preserving existing authentication, authorization, and audit boundaries without granting production write authority.The curator therefore requires no second integration path or model retraining.

B.1 Benchmark Construction

The evaluation uses CLBench database exploration and 90 adapted APEX consulting tasks across six document worlds, with controlled models, runs, and cost-aware metrics.

  • Benchmark Construction: 40 CLBench questions probe database discovery across schema changes, while adapted APEX provides 90 consulting tasks across six heterogeneous document worlds.CLBench includes hidden SQLite databases and a migration after question 20; APEX tasks span PDF, XLSX, DOCX, and PPTX files.
  • Models and Run Protocol: The primary studies use gpt-5.4 at xhigh reasoning effort, while the no-drift study evaluates Sonnet 4.6 and Opus 4.7.Memory conditions use intfloat/e5-base-v2 embeddings.
  • Models and Run Protocol: CLBench uses five independently shuffled paired runs per configuration; APEX uses five stateful and three stateless runs with shared seeded shuffles.Reported intervals are 95% Student-t intervals over run-level aggregates, except the no-drift study, which uses across-run standard deviation.
  • Metrics and Accounting: Strict pass requires every criterion to pass, and pass-discounted reward combines criterion completion with a task-agent tool-call penalty.For CLBench, the exploratory-query budget B is 15; APEX also reports a fractional-criteria diagnostic.
  • Metrics and Accounting: Reported metrics include pass rate, reward, task-agent tool calls, tokens, and task-agent response cost, excluding memory-management and harness-internal calls from tool counts.Learning-curve points are across-run means of running mean reward, with one-standard-deviation shaded bands.

C Per-World APEX Results

APEX results are organized by world using absolute reward and task-agent calls, while the curation pipeline distills completed traces and optionally probes the environment before memory updates.

  • Per-World APEX Results: Table 3 reports per-world absolute strict pass-discounted reward and benchmark-counted task-agent tool calls, with confidence intervals and highlighted best values.Each world is identified by difficulty and task count; memory and harness calls are excluded from tool-call totals.
  • Distiller Preprocessing: The distiller converts a completed rollout into a compact evidence packet without terminal feedback, memory tools, or authority to mutate durable records.It preserves goals, constraints, chronology, decisive action-observation pairs, reusable findings, failures, and unresolved assumptions.
  • Curator Agent: The curator receives the distilled packet, feedback, existing records, and optional raw-trace access, then maintains scoped, actionable memories through CRUD operations.Its record contract includes category, confidence, retrieval scope, and one concise lemma; unsupported or redundant candidates can be narrowed, corrected, merged, or deleted.
  • Environment-Probing Curation: Environment probing adds read-only world-tool access and a verification step while leaving the curator’s inputs, memory schema, CRUD policy, and stopping rule otherwise unchanged.Probes test candidate scope, counterexamples, stale mappings, preconditions, and whether shorter procedures reproduce the evidence.
  • Memory Record Analysis: Probe-backed records are directly executable, and matched CLBench examples reduced queries from 4, 9, and 8 to 1, 2, and 1.The examples show explicit tables, joins, filters, aggregation grain, and current schema rather than only warnings or tentative mappings.

E.2 CLBench Database Exploration

A representative CLBench question shows the stateless baseline failing after seven queries, while the benchmark answer is 267 and depends on the correct review-to-item relationship.

  • GHCP (No Memory): The stateless agent answered 188 instead of the accepted 267 after seven database queries, receiving zero reward.It added an incorrect category condition to a question about reviewed office products absent from attributes data.
  • Question Setup: The question asks for office products with at least one review but no representation in the attributes data.The hidden SQLite setting requires discovering relevant tables, joins, encodings, and conventions.
  • Trajectory: The stateless run inspected the database schema and explored data before issuing a final count query with review and attribute existence conditions.The recorded final query used items_g1, fdbk_g1, and attrs_g1 and consumed all seven allowed exploratory queries.

GHCP + Mem

GHCP + Mem retrieves reusable records for CLBench, but its representative run still explores broadly before reaching the correct answer; the supplied passages also contrast this with probing.

  • Retrieved Memories: The retrieved memory records encode product-domain mappings, review conventions, grouped-rating joins, and traps involving incorrect price or rating aggregations.Examples distinguish g1 office products, g2 electronics, and g3 musical instruments, and specify review-table aggregation conventions.
  • Representative CLBench Result: In the representative run, memory-supported exploration inspected schemas, examined review and item relationships, and computed the missing-attributes count from ref_id sets.The final computation reported 267, matching the accepted answer.
  • APEX Comparison: The selected APEX comparison reports memory transferring a computation recipe for a site-level revenue z-score task, with the environment-probing condition returning the same correct result using fewer calls.The supplied passage truncates the exact probing call count and result continuation.

GHCP — NO MEMORY

Without memory, the agent exhausted 96 tool calls and returned the incorrect Darcylis result. The trajectory lacked a valid decisive cross-company computation and relied on accessible Impact data rather than a combined cross-company attachment.

  • GHCP — NO MEMORY: 96 tool calls ended in failure, with the agent naming Darcylis and z = 1.29 instead of Lorexa and z = −1.60.The grader reports that both requested criteria were unmet.
  • GHCP — NO MEMORY: The baseline had no available memory and performed representative filesystem discovery before its unsuccessful computation.The recorded trace includes directory listing and filesystem discovery calls, with omitted calls summarized separately.
  • GHCP — NO MEMORY: The computation used five Impact sites and calculated site-level values from Impact revenue and headcount data, but did not establish a valid cross-company comparison.The recorded calculation produced Darcylis as the largest absolute z-score under its observed inputs.
  • GHCP — NO MEMORY: The response explicitly stated that no combined cross-company site-level operational attachment was located, limiting the result to accessible Impact files.It used sample standard deviation with Impact’s 2024 US revenue and an Impact site-level monthly headcount file.

GHCP + Mem

Memory enabled the agent to complete the benchmark correctly, while environment probing validated the relevant files and computation procedure with fewer calls. The probed configuration returned Lorexa with z = −1.60 in six calls.

  • GHCP + Mem: The memory-enabled baseline also succeeded in 11 calls, transferring a computation recipe and returning Lorexa with z = −1.60.Its configuration received primary and fractional rewards of 0.89.
  • GHCP + Mem: The retained procedure allocates annual US revenue across 2024 sites and 12 months, divides by site headcount, averages monthly ratios by site, and compares Impact sites.The memory schema specifies the Manufacturing_Site_Operations workbook and company PnL US sheets as inputs.
  • GHCP + Mem: The displayed evidence is selective rather than a full transcript, with omitted calls represented by accounting summaries and observations capped at 500 source characters.Lane headers report total task-agent tool calls, while some displayed traces show only representative calls.
  • GHCP + Mem: Six calls produced the correct Lorexa result with z = −1.60, satisfying both evaluation criteria.The probed configuration used a validated workbook map and received primary and fractional rewards of 0.94.
  • GHCP + Mem: Environment probing searched for the operational workbook and company PnL files before executing the decisive calculation.The probe found one manufacturing workbook and seven matching PnL files, while the trace records six top-level calls.
Loading 2609.11060v1…