AuthorsSusheel Suresh, Hazel Mak, Sahil Bhatnagar et al.
TL;DR
Grounding Agent Memory with environment-probing curation turns post-task memory into an environment-informed process, raising CLBench pass rate from 39% to 73% while halving task-agent cost.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Persistent memory built only from completed trajectories can preserve errors, overgeneralize partial evidence, and retain stale knowledge under environment drift.
On CLBench’s evolving databases, GHCP (No Memory) reaches only 39% pass rate and 8.60 total reward, forcing repeated rediscovery and wasted tool calls.
HOW IT WORKS
Grounding Agent Memory connects a task agent, curator agent, memory index, and trajectory distillation into a post-task memory pipeline with environment-probing curation.
Think of the memory index as long-term storage, the task agent as working RAM, and the curator agent as a background process that checks facts against the live environment.
This design lets Grounding Agent Memory validate and scope records with read-only tools, enabling correctness and drift repair that a plain context window or trajectory-only reflection cannot provide.
DIAGRAM
This diagram shows how Grounding Agent Memory separates task-time execution from post-task environment-probing curation with distinct tool boundaries.
DIAGRAM
This diagram shows how Grounding Agent Memory is evaluated across CLBench drift and APEX document worlds with different memory configurations.
PROCESS
Sequential task Si execution
Grounding Agent Memory runs the task agent with environment tools TEi and read-only memory_read from the memory index Mi−1 to solve Si.
Post-task memory curation
After closure and feedback gi, Grounding Agent Memory instantiates the curator agent with CRUD tools over the memory index.
Trajectory distillation
Grounding Agent Memory applies trajectory distillation Dψ to compress τi into di, preserving decisive observations and retrieved records for the curator agent.
Environment-probing curation
Grounding Agent Memory lets the curator agent propose–probe–commit records using read-only environment-probing curation tools before updating Mi.
KEY CONTRIBUTIONS
Diagnosis of trajectory-only generalization failure
Grounding Agent Memory shows that trajectory-only curation in systems like Mem0 and ReasoningBank cannot repair stale or incorrect records, leading to only 39% CLBench pass rate.
Deployment-compatible environment-probing curator agent
Grounding Agent Memory introduces environment-probing curation where a least-privilege curator agent uses read-only tools to validate, scope, and refresh records before writing.
Cost-aware evaluation on CLBench and APEX
Grounding Agent Memory achieves 73% CLBench pass rate and cuts task-agent cost from $3.38 to $1.68, while improving reward gain per dollar in five of six APEX worlds.
RESULTS
Pass (%)
73 %
+34 pp over GHCP (No Memory)
Total reward
22.60
+14.00 over GHCP (No Memory)
Queries per question
4.7
4.1 fewer than GHCP (No Memory)
Task-agent cost
$1.68
$1.70 cheaper than GHCP (No Memory)
On CLBench’s 40-question drift schedule, which tests database exploration under schema changes, Grounding Agent Memory substantially increases strict pass rate and pass-discounted reward while reducing SQL queries and task-agent USD cost.
BENCHMARK
On CLBench’s 40-question drift schedule, which tests database exploration under schema changes, Grounding Agent Memory substantially increases strict pass rate and pass-discounted reward while reducing SQL queries and task-agent USD cost.
BENCHMARK
Total reward on CLBench 40-question drift schedule (Equation 3).
KEY INSIGHT
Grounding Agent Memory with environment-probing curation raises CLBench pass rate to 73% while reducing task-agent cost from $3.38 to $1.68.
This is surprising because adding an extra curator agent and environment-probing curation might be expected to increase overall cost, not nearly halve the task-time spend.
WHY IT MATTERS
Grounding Agent Memory enables agents to maintain compact, validated procedural and factual memories that stay aligned with drifting enterprise environments.
Builders can now bolt environment-informed, auditable memory onto existing GHCP-like stacks without retraining or expanding task-time interfaces, improving reliability and efficiency across long task streams.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…