ERRAND: Budgeted Maintenance of Agent Memory
Beining Wu, Zihao Ding, Jun Huang
cs.AI
Quick verdict
ERRAND prices memory rechecks against a running wage; preregistered synthetic study supports claims strongly, though the world matches its assumed drift model.
Summary
Deployed LLM agents inherit a 'briefing' of knowledge items that were true at handover but silently go stale as the world drifts. ERRAND treats revalidation as a priced errand: beliefs over checkable atoms relax in closed form under a two-state drift model, a single-peaked value-of-information index prices each recheck, and a deadband gate funds one only when value per action clears a running wage (a dual estimate of the budget); task traffic yields free 'en-route receipts,' and repair versions items rather than deleting. The scheduler is closed-form arithmetic with no model calls (about 40 ms per step). On a scripted 2,000-step tool-use world with a frozen Qwen3-8B, at equal errand budgets (6/12/24 actions per 100 steps), ERRAND beats eager revalidation on the preregistered primary metric (ITT) by 8.0/10.0/5.2 percentage points, spends 11.0% of steps when uncapped versus eager's 70.7%, and the margin widens to 12.9 points where briefing coverage is thinnest. Ablations attribute the most to belief calibration (minus 13.0 points without it). Limitations: one synthetic environment whose drift follows exactly the assumed two-state Markov form, a small knowledge roster (26 items), only two Qwen backbones, and existing memory systems audited on paper but not run as baselines.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Revalidation should be priced as an action competing with task steps: funding a recheck only when its value per action clears the budget's running shadow wage dominates unpriced spending at equal budgets.
- At equal errand budgets (b=6/12/24 per 100 steps), ERRAND beats eager revalidation on the preregistered primary ITT caliber by 8.0/10.0/5.2pp and clears every non-oracle policy on ITT in every setting.
- Restraint wins: uncapped ERRAND self-limits to 11.0% of steps while uncapped eager revalidation spends 70.7% and still finishes 4.5pp behind capped ERRAND.
- The premium is a shadow price of long-tail knowledge: the conditional margin over eager widens from 2.3pp on fully covered steps to 12.9pp where briefing coverage is thinnest, and the ordering premium vanishes (−0.01pp pooled) where the budget does not bind (Proposition B.5, Table E.7).
Strengths
Preregistered analysis with mechanical verdicts
Appendix C freezes nine named estimands with interpretation bands before any full-wave run; 7/9 land in band, and the two misses are handled with preregistered fallbacks reported verbatim (strict dominance resolves to 1/3 caps and is scoped to the base cap; the calibration ladder's lower segment is left unresolved at +3.3pp [-4.1, +10.8]). This candor is rare on arXiv.
Gain is attributed to pricing, not spending
Figure 5b runs five policies at matched 10.6% spend: calibration buys +2.6pp over a blind-belief scheduler and pricing buys a further +10.3pp; Table E.6 shows the deployed index beats a Whittle closed-form allocator by 2.2pp at the tight cap b=6, where ordering should matter most.
Zero-model-call scheduler with audited overhead
Anti-gaming instrumentation
Honest ceiling and boundary reporting
Weaknesses
World-model congruence by construction
No deployed external baselines
Small scale and narrow backbone coverage
Thin statistical resolution in places
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ERRAND (arXiv 2609.29545, cs.AI) targets staleness in *handed-over* agent memory: a frozen LLM policy consults a briefing of consolidated items that were true at deployment while the world drifts silently. The reframing is economic — a recheck ('errand') is an action competing with task steps for the same budget, worth funding only when its value per action clears the budget's running shadow wage. The paper claims (1) closed forms for belief relaxation under two-state drift (Eq. 1), a single-peaked value-of-information index (Eq. 2), exact deadband width (Eq. 3) and a durability horizon (Eq. 4); (2) a zero-model-call scheduler that clears every non-oracle policy on preregistered calibers; (3) a regime map in which pricing matters where briefing coverage is thin and where real drift decouples from calendar priors. Scope statements are unusually explicit: exactness only in the single-atom regime, exogenous briefing substrate, frozen policy, and self-consolidated memory untested.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ERRAND in one sentence?
ERRAND is a scheduler for LLM-agent memory that treats re-checking a possibly-stale knowledge item as a priced action competing with task steps, funding a recheck only when its value per action clears the action budget's running shadow wage.
How much does ERRAND improve over baselines?
On a scripted 2,000-step tool-use world with a frozen Qwen3-8B and equal errand budgets (6/12/24 recheck-actions per 100 steps), it beats eager revalidation by 8.0/10.0/5.2 percentage points on the preregistered primary metric (task success over all errand-relevant steps); when given no budget cap it self-limits to 11.0% of steps while eager revalidation spends 70.7%.
Is code or data available for ERRAND?
The paper states an anonymized release package (scheduler library, instrument manifest with dispatch tables and drift schedules, briefing roster) accompanies the submission; the environment builds on the MIT-licensed AgentOdyssey harness with frozen Qwen3-8B/4B served via vLLM, and the authors claim no table cell was hand-entered.
What are the main limitations of the paper?
It is evaluated in a single synthetic environment whose drift follows exactly the assumed two-state Markov form, with a small roster (26 items), only two Qwen backbones, and existing memory-maintenance systems audited on paper (Table B.3) but never run as baselines; upkeep of self-consolidated (not handed-over) memory is untested.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt: '[...sampling gap...]' markers skip parts of Section 3.2 (errand grades), Appendix B, Appendix F and stretches of the reference list; conclusions about those portions rest on the surrounding intact text and tables.
- Figures are seen only through captions and embedded axis/table text; plotted data beyond the printed numbers cannot be verified.
- The reference list is partially sampled, so the bibliometric observation about self-citation density describes the visible portion only.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.