22
Useful Memories Become Faulty When Continuously Updated by LLMs
Editorial reading guide · 编者导读
为什么值得读
Agent memory 很容易把 compact、reusable 与 authoritative 混成同一个词。这篇给出具体 counterexamples，说明 continuously updated textual lessons 会删细节、添伪规则、跨任务覆盖；但 close read 也发现 headline fraction、raw-trajectory wording、Mind2Web scope、missing-outcome denominator 与 paper-to-release provenance 都需要收窄。把 paper、69-page PDF、August expanded release 与55,075 released rows分层后，真正留下的是一套 evidence-preserving consolidation contract，而不是一句 anti-summary slogan。
Source-backed paraphrase
论文在解决什么
这篇把 continual memory consolidation 从默认的 self-improvement primitive 改写成一种需要被证明安全的 lossy state transition：同一批 useful episodes 在 repeated LLM rewrite 后，utility 可以先升后降，schedule/grouping 也会改变 end state。它最重要的主张不是 raw memory 永远优于 abstraction，而是 episodic evidence 与 derived lesson 不能共享同一 authority；consolidation 应显式 gated、可回到来源、可撤销，并与 denominator、task pairing、schedule identity 一起被评价。
Source-backed paraphrase
关键抓手
- Memory objects：episodic traces保存单次发生了什么；consolidated abstractions跨 episodes提取 reusable lesson。论文测试 repeated textual rewrite 是否保持 utility，而不是把两者当作同一存储格式。（Abstract；§§1–2，PDF pp. 1–3）
- Non-monotonic utility：在 trajectories有用甚至来自 ground-truth solutions时，memory utility仍可先升、plateau、再下降到 no-memory baseline 以下；作者把损失归于 detail deletion、spurious rules、cross-task overwrite与 abstraction drift。（§§1、4–5；Figures 1–10，PDF pp. 2–8）
- Schedule as treatment：Static-All、streaming batches、order permutations与 repeated near-duplicates可从同一 episode collection形成不同 end states；update operator因此不能默认 exchangeable。（§§4.2–4.4；Figures 3、8–10；Appendix F，PDF pp. 5–7、44–52）
- Gated consolidation：controlled ARC-AGI Stream让 agent选择 Retain/Delete/Consolidate；auto regime偏向保留 episodes，并约为 forced-consolidation accuracy的两倍，episodic-only与 auto相当。它支持显式 action gate，不证明所有 agents都会学到正确 retention policy。（§§5–6；Figures 11–12，PDF pp. 8–10）
- Headline tension：19-task ARC treatment在同一v1中出现 failure 54%、failure 46%、accuracy 54%与Figure 2的10/19=52.6% accuracy四种不兼容写法；Static又被称为at the ceiling但画成18/19=94.7%。public prose必须回到exact fraction与locator。
- Comparator wording：AppWorld所谓raw trajectory control实际使用Qwen3.5-27B对完整ReAct trajectory的episode-level prose summaries；安全名称是episode-level summarized records with no cross-trajectory lesson abstraction，而不是byte-raw histories everywhere。（Tables 2–3；Appendix F.4，PDF pp. 7、15、48–49）
Source-backed argument map
论证怎样成立
01 · 先区分 episode evidence 与 derived abstraction
作者把 append-only/retained episode records 与 repeatedly rewritten textual lesson bank设成不同 memory forms，避免只用最终QA score把 evidence retention与abstraction utility混在一起。
Abstract；§§1–3，PDF pp. 1–4
02 · 在useful-input条件下寻找consolidation regression
ScienceWorld、ALFWorld、WebShop、AppWorld与ARC settings刻意使用agent已解或ground-truth experiences，使 observed regression更难被输入噪声单独解释。
§§3–4；Figures 1–3；Tables 1–3，PDF pp. 3–7
03 · 用size/checkpoint curves展示utility非单调
多种methods在memory增长过程中出现early gain、plateau或later decline；作者据此反驳更多consolidated experience必然更有用的monotonic assumption。
§4；Figures 1、4–10，PDF pp. 2、5–7、32–47
04 · 用schedule与memory surgery提出机制
同一pool的order/grouping会改变memory end state，删除/替换selected harmful content有时恢复performance；这些是mechanism clues，不是prevalence或significance estimate。
§§4.2–4.4、7；Appendix F、I.1，PDF pp. 5–7、10、44–52、61–62
05 · 让agent选择memory action
ARC Stream把Retain/Delete/Consolidate暴露为actions；auto与episodic-only优于forced consolidation，使‘每轮必压缩’从默认流程降为可比较treatment。
§§5–6；Figures 11–12，PDF pp. 8–10
06 · 把practical recommendation收窄为evidence-first gate
作者建议raw episodes作为first-class evidence、consolidation显式gated；paper limitations又把结论限制在textual abstractions、current models、few repeats与point estimates。
§§8–9；Limitations，PDF pp. 10–11
Source-backed method notes
方法与监督
Cross-environment descriptive surfaces
paper组合ScienceWorld、ALFWorld、WebShop、AppWorld与ARC-AGI/controlled Stream；method、model、task count、trajectory representation与metric不同，因此跨panel共同支持failure mode existence，不形成统一effect size或system ranking。
§§3–6；Tables 1–3；Figures 1–12，PDF pp. 3–10
Ground-truth ARC isolation
先筛选no-memory可解的19个ARC tasks，再以ground-truth solutions形成memory，试图把input-side trajectory error与consolidation error分开；v1 headline数字内部冲突，因此只能按exact fraction/figure逐项署名。
§4.1；Figures 2–3；Appendix C，PDF pp. 3、5、23–31
Controlled memory actions
ARC Stream显式提供Retain、Delete、Consolidate，并比较auto、forced-consolidation与episodic-only regimes；它观察policy/action-space interaction，不等于production agent的owner authority或长期retention governance。
§§5–6；Figures 11–12，PDF pp. 8–10
Paper-reported findings
报告了什么
- paper-reported：ScienceWorld中若干methods随memory size先改善后下降；WebShop AWM随examples增加而success下降，episode-level summarized control更稳。Figure 1同时展示binary success与parenthesized average score，不能把两种metric互换。（Figure 1，PDF p. 2）
- paper-reported：controlled ARC Stream里auto与episodic-only约为forced consolidation accuracy的两倍；作者观察auto agent多数保留episodes，而非持续consolidate。（§§5–6；Figures 11–12，PDF pp. 8–10）
- paper-reported：WebShop W8 memory surgery在50 tasks的single run中删除/替换selected content后performance变化；作者明确没有significance test，因此属于mechanism hypothesis与case evidence。（§7；Appendix I.1，PDF pp. 10、61–62）
AMS audit · 本站复核与实验
AMS 复核与实验
以下为本站对固定代码、公开记录或原创场景的检查；适用范围以所链接的方法与记录为准。
检查对象与观察
- Exact-current audit：762个arm summaries与55,075 released rows全部内部一致；official 675 numeric assertions只resolve到335 arms，304 missing outcomes影响40 arms，selected current/A/B/succdrop pools保持同一200-record canonical-JSON multiset但historical run lineage未被digest封存。（public checked artifact）
本站已记录的结果
- released-row audit：exact current release有762 arms、55,075 rows、54,771 valid outcomes与304 invalid/refusal/skipped rows；所有arm sets、row identities、metadata、count summaries与ARC grids内部一致。（public raw/audit.json）
- official-scope audit：upstream 675 assertions PASS，但只resolve到335/762 arms；这些arms关联33,956/55,075 raw rows。40张number tables中27张、511 data rows未被mapper打开，arms.csv也不在official numeric verifier中。（public raw/audit.json）
- schedule fixture audit：natural/A/B/succdrop各200 records且full canonical-record multiset相同；A/B/succdrop移动196/199/199个positions。N25的B result arm只有49 task identities，N50/N100/N150/N200四schedules才完全paired。（public raw/audit.json）
展开复核方法与推论
Current-release membership and denominator audit
standard-library checker独立读取六个arms.csv与merged JSONL，按environment-specific row identity、Boolean success、metadata、aggregate与ARC rectangular grid重算762 arms/55,075 rows，并同时保留task-denominator与valid-only rates。
research/faulty-memory-release-boundary-audit/PROTOCOL.md；REVIEW-AMENDMENT.md；REVIEW-AMENDMENT-2.md；raw/audit.json
Official-verifier coverage instrumentation
先运行exact upstream verifier，再对锁定的13张mapper tables独立枚举675 assertions、requested/actual arm与fallback；coverage unit分别报告assertions、335 unique arms、associated raw/valid rows以及27张unmapped tables。
official results/verify_numbers.py；public raw/audit.json
Schedule fixture and receipt-chain validation
selected pools比较完整canonical JSON objects与ordered IDs，exceptions单列；final checker用17 negative/2 positive production controls、两个fresh seeds、exact section/whole-audit digests、exact JSON types、canonical receipt bytes与no-extra-file inventory区分current source-bound revalidation、producer-recorded comparison和offline package consistency。
research/faulty-memory-release-boundary-audit/audit.py；raw/mutation-controls.json；raw/repeatability.json
Evidence boundary
仍未证明什么
- 论文自己限定text-only、natural-language abstraction、current model capability、few repeats、point estimates与缺少formal error bars；embodied、multimodal、structured与parametric memory不在证据范围。（Limitations，PDF p. 11）
- WebShop W8 surgery是n=50、single run、no significance test；case recovery不能估计failure prevalence、跨task effect或universal causal law。（Appendix I.1，PDF pp. 61–62）
- paper panels使用不同models、methods、task sizes、metrics与trajectory representations；跨environment相似curve支持counterexample family，不支持一个pooled consolidation effect size。
- episode control在AppWorld是model-generated prose summary而非raw token trajectory；representation、retrieval/context budget与cross-episode abstraction同时变化，不能只用raw-vs-summary二元标签解释结果。
- August public source晚于May paper且没有production-revision binding；current released-row consistency不能选择paper headline真值、恢复historical run或替代paper experiment reproduction。
- selected pool bytes相同不证明historical result arm实际消费该pool；arms.csv没有pool digest，缺少model/environment/request receipt与runtime lineage。
- 304 missing outcomes改变40个arms的observed level，部分ScienceWorld curve direction也随denominator改变；invalid/refusal不是自动observed failure，task coverage与conditional-valid performance必须并列。
- 17 negative/2 positive controls测试AMS checker。stored comparison只验证producer-recorded internal bindings，不自证两次historical processes；fresh source-bound invocation才在当次调用中执行两棵roots并匹配stable bytes。Checksums不是external signature，receipt-only不重新打开source/PDF，也不证明cross-platform reproducibility或security。
Source audit · observation and consequence kept distinct
原文内部张力
19-task headline有四种互斥读法
Abstract说fails on 54%，Introduction说fails on 46%，§4.1说accuracy down to 54%，Figure 2画Stream R10为10/19=52.6% accuracy，即9/19=47.4% failure；Static同时被称at the ceiling但图上是18/19=94.7%。
定位：Abstract，PDF p. 1 · Introduction，PDF p. 2 · Figure 2，PDF p. 3 · §4.1，PDF p. 5
编者 consequence：public prose应给exact fraction与locator，并把Static写成near the ceiling。current release不能反向决定May historical run的正确headline。
Raw trajectory control实际是episode summary
Table 2 heading写Trajectory Logs (Summarized)，caption又称raw；Table 3使用summarized rollouts；Appendix F.4说明AppWorld datapoint是Qwen3.5-27B对full ReAct trajectory的prose summary。
定位：Tables 2–3，PDF pp. 7、15 · Appendix F.4，PDF pp. 48–49
编者 consequence：control只排除cross-trajectory lesson abstraction，不排除episode内compression或summary-model error；rawness、abstraction level与context budget应拆开。
Mind2Web点名但没有result surface
§3与Limitations点名Mind2Web，却没有setting row、result table、figure或appendix experiment；current release转而宣传PropWorld与ARC-AGI。
定位：§3，PDF p. 4 · Limitations，PDF p. 11 · exact current source README/CITATION.cff
编者 consequence：Mind2Web保持paper-scope unresolved；PropWorld是later release surface，不是缺失paper result的replacement或synonym。
Official 675 cells不是full-release verification
upstream verifier的cell是13 selected CSV mappers产生的numeric assertion；675 assertions含重复arm checks并经两个allowlisted ALFWorld fallbacks只resolve到335 unique arms。27/40 number tables与arms.csv不进入该path。
定位：official results/verify_numbers.py · public raw/audit.json
编者 consequence：675只能描述specific arithmetic surface。完整release membership、denominator、row identity与unmapped tables需要独立gate，不能从0 mismatched外推。
Paper与current release不是同一denominator object
paper setting counts与current releasedarm rows多处不同；current interactive release还有304 missing outcomes。ScienceWorld mem0在task denominator下单调下降，在valid-only denominator下非单调。
定位：paper settings/appendices · public raw/audit.json denominators/registered_curves
编者 consequence：curve claim必须绑定source revision、attempted/valid denominator与paired identity sets；current released curves不替代paper-time estimand。
Editorial synthesis / question
编者判断
Consolidation is a candidate, not evidence authority · Editorial inference — not tested by the paper
跨episodes抽出的lesson是对evidence的可反驳解释。它应有独立identity、support links、schedule/prompt/model receipt与rollback，而不应静默覆盖supporting episodes。
这是paper failure cases与evidence-preservation recommendation的architecture synthesis；不是论文已实现的promotion protocol或对任何production system的failure判定。
Schedule belongs in memory identity · Editorial inference — not tested by the paper
若同一input multiset因order/batching产生不同lesson，schedule不是无关execution detail，而是derived state provenance的一部分；comparison必须content-lock pool并seal treatment lineage。
selected shipped fixtures证明content conservation，未证明historical runs的pool consumption或causal order effect。
Denominator observability is memory observability · Editorial inference — not tested by the paper
refusal/error/skipped改变的不只sample size，还会改变curve shape。attempted-task coverage与conditional-valid utility回答不同问题，应同时进入memory evaluation receipt。
current release提供一个exact denominator sensitivity case；它不宣布某个universal missing-data estimator或对paper historical runs重新计分。
Episodic-first does not mean context-dump everything · Editorial inference — not tested by the paper
first-class episode evidence可以存于durable、可定位层，由retrieval按需进入active context；保留evidence与无限token replay是两个设计决定。
paper比较若干episode controls与consolidators，但没有验证一种production-gradestorage/retrieval architecture或最优context budget。
A pass can coexist with a revise verdict · Editorial inference — not tested by the paper
current released rows可以在内部完全一致，同时paper/public wording仍因headline、representation与provenance tension需要revision。numeric gate与editorial claim gate保护的是不同truth layers。
PASS只属于exact-current released-row contract；它不提升paper-time identity、causal inference或semantic outcome truth。
继续追问
如果同一批useful episodes只因update order或batch grouping不同就生成不同lesson，consolidated memory凭什么覆盖它所依赖的evidence，而不是先作为一份可撤销、可反驳、带schedule lineage的candidate存在？
继续追问与实践
经验在什么条件下能够提高当前能力？环境、版本或适用范围改变以后，怎样继续受益，又不被过去误导？
持续研究 · 当前判断可修订
工具参数、配置 schema 或工作经验随版本改变，但旧环境仍然在使用；或多个来源对同一范围给出不同建议时。
编者建议 · 目标侧待验证
Editorial transfer · 编者设计建议
读完，可以怎样借用
基于既有精读与本条明确标注的证据范围整理；下面针对新场景的迁移对照尚未执行。
- 什么时候有用
- 系统反复把经历重写成经验，希望记忆越积累越有用。
- 可以借哪一步
- 把派生 lesson 作为可撤回的候选保留，同时保存其 supporting episodes 和更新顺序；在实际使用时说明适用条件。
- 先做什么对照
- 固定同一组公开经历，比较一次性、逐批和反序归纳，保留 episode 对照；再更正来源，看 lesson 是否仍可解释。分别记录完成、失败与缺失。
- 带走时的边界
- 这组模型迁移对照未执行。已有 released-row audit 与专题的规则演示都不能证明真实 consolidation 的因果效果，也不支持“原始记录永远更好”。
阅读依据：既有札记 §§4–6、Appendix F.4、公开测试；“编者判断”：candidate / schedule / denominator。
Editorial protocol · proposed, not run
可复核的下一步
Proposed · not run
Evidence-preserving consolidation candidate protocol
当Static-All、fixed-order stream、reversed stream、shuffled batches与episodic-only treatments共享exact episode multiset时，derived lesson能否在保留claim-to-episode lineage与owner correction authority的同时，提高held-out utility，并对schedule变化保持可解释稳定？
- Method
- Owner-seal 24–48组public/synthetic episode bundles与atomic support/obsolete/conflict labels；content-address每个input multiset与schedule；consolidator只生成独立candidate，不写canonical state；由blind human/rule-grounded adjudication分别评价support preservation、unsupported addition、obsolete resurrection与legitimate abstraction，再用matched context budget测held-out task utility。
- Public / synthetic fixtures
- exact episode IDs/revisions；support facts；legitimate correction；obsolete fact；cross-episode pattern；contradictory/ambiguous controls；Static-All、forward/reverse/shuffled streams；episodic-only与lossless extractive-index controls；candidate claim-to-source links与raw request/response receipts。
- Controls
- source/model/prompt/schedule/context-budget hashes；same-multiset pairing；generator与judge separation；owner-authority cases；no-memory and episodic controls；forced unsupported-addition/deletion mutations；abstention allowed；three-run schedule/model repeat；candidate never receivescanonical write authority。
- Measures
- support-preservation recall；unsupported-addition and obsolete-resurrection rates；legitimate-abstraction false rejection；claim-level source recoverability；cross-schedule semantic variance；held-out utility with attempted/valid denominators；compression ratio；abstention；cost/latency/egress separately。
- Limitations
- 这是proposed-not-run protocol，需要另行sealed gold、model/provider/privacy/budget与owner acceptance。它不会恢复paper historical runs，也不授权external model call、Tilia mutation、production promotion或从small synthetic bundles外推real-world prevalence。
Attributed public perspectives and tests
署名观点与测试
Exact-current released-row, verifier-coverage, denominator and schedule-fixture audit
- Method
- Reject a wrong/dirty checkout or wrong paper; discover and lock the six exact experiment surfaces, source tree/origin, selected input blobs and paper digest; independently validate environment-specific arm/row membership, Boolean outcomes, metadata, aggregates and ARC grids; preserve task and valid denominators; execute then instrument the upstream mapper for assertion/arm/row/table coverage; compare full canonical JSON pool multisets and registered exceptions; build exact ordered curve registries and shared-checkpoint pairing receipts; execute 17 production-path negative and two positive controls; run two fresh roots with distinct hash/traversal seeds; bind checker/source/paper/stable-file hashes through compare/install; require exact JSON types, canonical receipt bytes, whole-audit digest and exact no-extra-file inventory; verify package integrity separately from fresh source-bound stable-byte matching.
- Environment
- Darwin arm64；CPython standard library；exact clean official commit adf347f67a81a19fc71d2529d09108c71a1f9712 and exact 69-page arXiv v1 PDF；fresh external-storage roots；PYTHONHASHSEED/traversal seeds 313 and 727；no model, API, agent environment, benchmark runtime, credential, private data or network-dependent experiment called。
- Fixture
- Six arms.csv files；six merged per-task JSONL files；762 arms；55,075 rows；54,771 valid and 304 invalid/refusal/skipped outcomes；13 official-mapped and 27 unmapped number tables；natural/A/B/succdrop selected 200-record pools plus D/drop50/replay2 exceptions；seven ScienceWorld curves and four AppWorld schedule curves；exact paper/source manifests；two fresh run roots and installed public receipts。
- Controls
- Exact commit/tree/origin/paper/input locks；actual experiment discovery；clean pre/post status；unique environment-specific row identities；all-row Boolean success；metadata/count/grid recomputation；675-assertion and 259 mapped-row freeze；fallback allowlist；assertion-versus-arm unit validator；missing-outcome semantics；non-empty pool IDs and full-object multiset equality；exact curve key/checkpoint registry and CSV-order invariance；17 negative plus two positive production controls；stable section and whole-audit digests；exact JSON scalar types and canonical raw receipt bytes；checker/run/comparison/install bindings；exact eleven-file checksums with no ignored extras；bounded public package path/credential/payload-key scan；adversarial type/serialization/file-set regression fixtures。
- Raw result
- Final formal receipts report PASS for 762 arms and 55,075 rows, including 54,771 valid outcomes and 304 invalid/refusal/skipped rows across 40 affected interactive arms. The exact upstream verifier reproduced 675 assertions with zero mismatches/no-row cases; instrumentation resolved them to 335 unique arms, 33,956 associated raw rows and 33,658 associated valid rows, while 27/40 number tables containing 511 data rows were unmapped. The maximum denominator delta was ScienceWorld eval/letta_N400 at 20.25 percentage points. Selected A/B/succdrop pools preserved the natural 200-record canonical-JSON multiset while moving 196/199/199 positions; pool_D/drop50/replay2 remained explicit exceptions and result-to-pool cryptographic run lineage was false. N25 B used a distinct 49-task set; the four schedules were paired at shared N50/N100/N150/N200, while B-only N85 was not cross-schedule comparable. In the recorded local procedure both fresh runs detected 17/17 negative controls, passed 2/2 positive controls and produced two byte-identical stable receipts bound to exact checker/source/paper metadata; the stored comparison labels historical process identity PROCEDURAL_NOT_RECEIPT_PROVEN. Receipt-only package consistency and a separate fresh source-bound two-root stable-byte match passed. These are executed checked-receipt results, not paper-reported numbers or reconnaissance claims.
- Derived result
- At exact current public commit adf347f67a81a19fc71d2529d09108c71a1f9712, all 762 released arm summaries are internally consistent with 55,075 released rows under the registered identity, missingness and aggregate rules. The official 675 assertions are a partial arithmetic surface reaching 335 arms, not complete-release verification. The selected current/A/B/succdrop shipped pool files conserve one 200-record full-content multiset while changing order, but their historical result-to-pool binding and causal schedule effects remain unestablished. The audit therefore supports released-row consistency PASS and narrower paper/public wording REVISE at separate evidence layers.
- Limitations
- The protocol was frozen after full paper/source reconnaissance and official gates, so it is registered relative to independent checker execution but not blinded. One review rejected preliminary receipts; REVIEW-AMENDMENT.md was frozen before corrected execution. A second adversarial review rejected those corrected receipts before installation; REVIEW-AMENDMENT-2.md was frozen before final execution, and only final receipts are installed. Membership/aggregate agreement does not prove stored-outcome semantics or historical environment behavior. Missing paper-time revision/run/model/pool bindings remain missing. Synthetic controls test the checker. Stored run labels, seeds and environment hashes are producer metadata rather than execution attestation; same-machine byte identity and local package digests are not external signatures, cross-platform reproduction or correctness proofs. Source code/data licensing and third-party benchmark/model terms remain distinct; this AMS-authored artifact receives no new license grant pending owner choice.
Evidence basis：Paper-facing audit of arXiv:2605.12978v1 and exact clean official DylanZSZ/Memory-Collapse-Eval commit adf347f67a81a19fc71d2529d09108c71a1f9712, covering all six released arms.csv/merged JSONL surfaces, all 762 arms and 55,075 rows, the upstream 675-assertion verifier, 40 number tables, selected AppWorld pool fixtures and registered ScienceWorld/AppWorld curves.
Boundary：This is an independently authored, deterministic, model-free exact-current-release consistency audit under a frozen non-blinded protocol and two explicit chronology-preserving post-review amendments. It is not a paper experiment reproduction, paper-time source identification, model/agent/environment/benchmark run, semantic outcome validation, historical execution replay, causal schedule estimate, significance test, system ranking, prevalence/harm estimate, security audit, production reliability claim or real-user generalization. The stored comparison validates producer-recorded internal bindings but does not attest historical process identity. The evidence artifact does not itself assign canonical AMS reading depth.
method, results and claim ceiling ↗ · frozen original protocol ↗ · frozen post-review amendment ↗ · frozen second post-review amendment ↗ · source-locked corrected checker ↗ · complete stable audit receipt ↗ · production-path control receipt ↗ · producer-recorded stable-byte comparison ↗ · receipt-only and source-bound verifier ↗
Shared readings
放到一起读
一条更正之后 →
派生经验不能凭“总结过了”取得覆盖其原始证据的权威。
论文的 model-generated consolidation 与这里的 typed rule 更新不是同一 treatment。
串读定位：§4.1、Tables 2–3、Appendix F.4；见材料的 source tensions
含当前正文、出处、阅读范围与接续线索；链接的全文不随包附带。