AuthorsHaoyu Wang, Guangyuan Dong, He Liang et al.
TL;DR
MemGuard persists verifier descriptors across the memory lifecycle, yielding up to +7.9 SR points on WebArena over ReasoningBank.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM agents suffer unreliable admission and memory drift, where failed trajectories and stale records enter memory and mislead later decisions.
Terminal, web, and software-engineering agents then reuse misleading observations and outdated commands, causing wrong edits, incorrect navigation, and degraded reliability over long task streams.
HOW IT WORKS
MemGuard combines Verifier-Guided Admission, Structured Memory Units, Precision-Oriented Retrieval, and Memory Consolidation and Budgeted Governance to control which experiences become reusable memory.
Think of MemGuard as a smart card catalog plus librarian: every memory card stores verifier scores and lifecycle state, and the librarian continuously re-files, merges, or retires cards.
This persistent governance lets MemGuard reject risky experiences, keep provisional records, and retire stale or conflicting memories in ways a plain context window or append-only memory bank cannot.
DIAGRAM
This diagram shows how MemGuard processes each task using Algorithm 1, from retrieval through verification to memory update.
DIAGRAM
This diagram shows how MemGuard consolidates, resolves, summarizes, and archives memories under a fixed active-memory budget.
PROCESS
Precision-Oriented Retrieval
MemGuard computes S+(xt, m) and S−(xt, m) to retrieve positive memories and failure guards from the governed memory bank, prioritizing relevance and verifier quality.
Verifier-Guided Memory Admission
After the agent produces trajectory τt and output yt, MemGuard scores completion, consistency, validity, and generalizability to derive Rt, ct, ℓt, and uncertainty ut.
Structured Memory Units
MemGuard’s induction module converts trajectories into structured records with titles, content, lifecycle state, usage stats, conflict links, and attached descriptor dm = (Rm, cm, ℓm, νm).
Memory Consolidation and Budgeted Governance
MemGuard merges duplicates, resolves conflicts using Eq. 5, summarizes repeated successes, archives stale records, and enforces the active-memory budget B.
KEY CONTRIBUTIONS
Verifier-Guided Admission
MemGuard decomposes trajectory verification into multiple criteria and attaches dm = (Rm, cm, ℓm, νm) to each memory, rejecting or making provisional any candidate below Rt ≥ 0.70 and ct ≥ 0.60.
Structured Memory Governance
MemGuard represents experiences as structured records with lifecycle state, usage statistics, deduplication signatures, and conflict links, turning the memory bank into a lifecycle-managed object instead of an append-only index.
Comprehensive Evaluation Across 16 Settings
MemGuard is evaluated on Terminal-Bench 2.0, SWE-Bench Verified, WebArena, and Mind2Web across four backbones, achieving the best success metric and lowest average steps in all 16 backbone–benchmark settings.
RESULTS
WebArena SR
58.4%
+7.9 points over ReasoningBank
Mind2Web SSR
51.8%
+5.6 points over ReasoningBank
Terminal-Bench 2.0 SR
67.4%
+2.8 points over ReasoningBank
SWE-Bench Verified Resolve Rate
83.6%
+3.5 points over ReasoningBank
On Qwen-3.5-Plus, WebArena and Mind2Web test web navigation and multi-step web tasks where unstable states easily corrupt memory. These gains show that MemGuard’s persistent verifier governance materially improves long-horizon agent reliability over ReasoningBank.
BENCHMARK
On Qwen-3.5-Plus, WebArena and Mind2Web test web navigation and multi-step web tasks where unstable states easily corrupt memory. These gains show that MemGuard’s persistent verifier governance materially improves long-horizon agent reliability over ReasoningBank.
BENCHMARK
Success Rate (SR) on WebArena comparing MemGuard against memory baselines.
KEY INSIGHT
The verifier-only filter already improves WebArena SR to 52.0%, yet MemGuard reaches 58.4%, adding another 6.4 points while also reducing average steps from 8.4 to 6.9.
This is surprising because one might expect better one-shot verification to capture most benefits, but MemGuard shows that persisting verifier signals as lifecycle metadata yields substantial extra gains.
WHY IT MATTERS
MemGuard unlocks governed long-term experience memory where every record carries persistent verifier-derived descriptors that shape admission, retrieval, conflict resolution, and archival.
Builders can now deploy agents that accumulate experience across hundreds of tasks without uncontrolled memory drift, using MemGuard to keep banks compact, reliable, and resistant to misleading trajectories.
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…