When Retrieval Is Not Enough: Obsolete Facts in Agent Memory

# When Retrieval Is Not Enough: Obsolete Facts in Agent Memory
Source: [https://arxiv.org/html/2610.36308](https://arxiv.org/html/2610.36308)
\\workshoptitle

Personalized, Aligned, Long\-Term Memory for AI Systems

Oleeviya Babu PoikarayilAffiliation:Faculty of Computer Science, Otto\-von\-Guericke\-Universität Magdeburg, GermanyEmail:[oleeviyazehr@gmail\.com](mailto:)Abdulrahman NahhasAffiliation:Otto\-von\-Guericke\-Universität Magdeburg, GermanyEmail:[abdulrahman\.nahhas@ovgu\.de](mailto:)Het Darshan MehtaAffiliation:Otto\-von\-Guericke\-Universität Magdeburg, GermanyEmail:[het\.mehta@ovgu\.de](mailto:)Jawid Ahmad BaktashAffiliation:Technical University of Munich, GermanyEmail:[jawid\.baktash@tum\.de](mailto:)Ernesto William De LucaAffiliation:Otto\-von\-Guericke\-Universität Magdeburg, GermanyEmail:[ernesto\.deluca@ovgu\.de](mailto:)Klaus TurowskiAffiliation:Otto\-von\-Guericke\-Universität Magdeburg, GermanyEmail:[klaus\.turowski@ovgu\.de](mailto:)

###### Abstract

Memory\-agent benchmarks often collapse retrieval and answer generation into one score, making it hard to tell whether a failure comes from missing an update or from not using an update that was already retrieved\. We examine this distinction on the single\-hop FactConsolidation task of MemoryAgentBench\. With Qwen2\.5\-7B\-Instruct on 6K\-history BM25 accuracy drops from 45% to 32% when the retrieved evidence budget grows from 64 to 512 tokens, even though a gold\-assisted post\-hoc audit finds the newest statement in every retrieved set\. Most of the errors occur on questions that also contain an obsolete statement version\. Under a frozen prompt and equal four\-statement evidence size, replacing one unrelated statement with the obsolete competing statement reduces accuracy from 100% to 44\.6% for Qwen2\.5 and from 100% to 87\.8% for Qwen3\-14B, with all 50 harmful paired flips classified as stale\. A 32K\-history replication likewise favors BM25\-64 over BM25\-512 \(69% versus 58%\)\. These results suggest that retrieving the correct fact is not enough\. An obsolete competing statement can be more harmful than an unrelated piece of evidence of the same size\.

## 1Introduction

Persistent memory lets an agent carry preferences, task state, and factual updates across sessions\. A deployed system must therefore solve at least two distinct problems: retrieving the relevant information, and selecting the currently valid version when memory contains both an old fact and a later correction\. End\-to\-end accuracy hides this distinction\. The retriever may return the newest fact, yet the language model can still answer with the older one\.

We ask the following question:*when the correct update is retrieved, how do evidence budget, unrelated distractors, and an older competing fact affect whether an LLM uses it?*We study this question using the MemoryAgentBench’s FactConsolidation task\([Hu et al\., 2026](https://arxiv.org/html/2610.36308#bib.bib1)\)\. In this task, numbered statements are chronological and higher numbers replace lower conflicting statements\. The facts are synthetic and can therefore contradict ordinary world knowledge\. For example, the memory may first state that the author of*The Marriage of Figaro*is Pierre Beaumarchais and later state that it is Thomas Kyd\. Under the benchmark’s chronology, Thomas Kyd is then the correct answer\.

We address the question in three parts:RQ1asks how the evidence budget affects the use of a retrieved update;RQ2asks whether failures are concentrated on questions for which an obsolete value is also retrieved; andRQ3asks whether an obsolete competing fact is more harmful than unrelated evidence of the same size\.

We vary the BM25 budget, compare questions with and without an obsolete competitor, and run an equal\-size substitution on two models, with a 32K budget replication\. Together, these test a focused claim:obsolete target evidence is substantially more damaging than unrelated evidence of the same size\.Appendix[B](https://arxiv.org/html/2610.36308#A2)provides a complete inventory of the exploratory, diagnostic, final, replication, and excluded runs\.

#### Relation to prior work\.

Memory benchmarks such as LoCoMo, LongMemEval, and MemoryAgentBench cover multi\-session interaction and temporal updates\([Maharana et al\., 2024](https://arxiv.org/html/2610.36308#bib.bib3);[Wu et al\., 2025](https://arxiv.org/html/2610.36308#bib.bib2);[Hu et al\., 2026](https://arxiv.org/html/2610.36308#bib.bib1)\), while RAG/memory systems provide compact evidence\([Lewis et al\., 2020](https://arxiv.org/html/2610.36308#bib.bib4);[Chhikara et al\., 2025](https://arxiv.org/html/2610.36308#bib.bib8)\)\. Long\-context studies document distraction and contradictory\-evidence effects\([Shi et al\., 2023](https://arxiv.org/html/2610.36308#bib.bib6);[Liu et al\., 2024](https://arxiv.org/html/2610.36308#bib.bib7);[Kurfali and Östling, 2025](https://arxiv.org/html/2610.36308#bib.bib11)\); recent work also targets temporal or invalidated knowledge\([Wallat et al\., 2026](https://arxiv.org/html/2610.36308#bib.bib12);[Chao et al\., 2026](https://arxiv.org/html/2610.36308#bib.bib13)\)\. Most closely,[Reddy and Challaram \(2026\)](https://arxiv.org/html/2610.36308#bib.bib14)study FactConsolidation but note that their pipeline comparison jointly changes resolver, prompt, format, and temperature\. We extend their pipeline\-level finding with a controlled intervention\. It isolates evidence composition while holding reader, prompt, latest fact, and four\-statement size fixed\.

## 2Experimental design

#### Data and chains\.

We use theConflict\_Resolutionsplit of MemoryAgentBench, focusing onfactconsolidation\_sh\_6kandfactconsolidation\_sh\_32k\. The 6K history contains 455 statements and 100 questions\. Template\-based chain reconstruction resolves all 100 questions: 74 have one older conflicting statement and 26 have no previous conflict\. The 32K history contains 2,310 statements and 100 questions\. We use the benchmark’s normalized substring metric for the primary results and manually inspect semantic scoring edge cases\.

#### Retrieval budgets\.

We rank individual statements with BM25\([Robertson and Zaragoza, 2009](https://arxiv.org/html/2610.36308#bib.bib5)\)\. We add them until reaching a token budget, and restore chronological order\. The 6K sweep uses budget of 64, 128, 256, and 512 evidence tokens; the 32K replication uses the two endpoints\. A post\-hoc support audit uses the benchmark answer only to check whether the intended latest statement is present in the retrieved evidence\. It is therefore a diagnostic, not a deployable retrieval procedure\.

#### Equal\-size intervention\.

For each of the 74 conflict questions, we build two four\-statement contexts: \(A\) the latest target statement and three unrelated statements, and \(B\) the obsolete target statement, the latest statement, and two of the same unrelated statements\. The distractors are selected deterministically, have zero content\-word overlap with the question, and exclude update\-chain values\. Thus A and B differ only by replacing one unrelated statement with the obsolete competitor\. These 74 questions were also used for prompt development and the pilot\. This is a within\-cohort paired comparison rather than held\-out confirmation\. Freezing the prompt across conditions protects the A/B contrast\.

#### Models and inference\.

We evaluateqwen2\.5:7b\-instructandqwen3:14b\([Yang et al\., 2024](https://arxiv.org/html/2610.36308#bib.bib9);[Team, 2025](https://arxiv.org/html/2610.36308#bib.bib10)\), served through Ollama on a university GPU system\. Local Python scripts send requests to the HTTP chat API over a VPN\. The final controlled experiments and corrected 32K replication use temperature 0, seed 42,num\_ctx=2048, andnum\_predict=64; the Qwen3 run additionally setsthink=false\. Earlier 6K runs use larger context settings \(Appendix[A](https://arxiv.org/html/2610.36308#A1)\)\. The frozen final prompt instructs to use only supplied numbered memories, treat higher\-numbered conflicting statements as newer replacements, and return the shortest answer phrase\. We use paired exact McNemar tests as benchmark diagnostics; questions share a history, so we avoid population\-level interpretations\.

## 3Results

### 3\.1Budget growth exposes a conflict\-resolution failure

On the 6K history, BM25 accuracy decreases monotonically from 45% at 64 tokens to 32% at 512 tokens \(Table[1](https://arxiv.org/html/2610.36308#S3.T1)\)\. A gold\-assisted post\-hoc audit identifies the latest statement in all 100 retrieved sets at every budget\. The decline in accuracy therefore cannot be explained simply by loss of the intended update\. In the paired 64\-versus\-512 comparison, 14 questions are correct only at 64 tokens, while just one is correct only at 512 \(p=9\.77×10−4p=9\.77\\times 10^\{\-4\}\)\.

Table 1:Evidence\-budget results\. “Conflict” uses reconstructed chains\. Gold\-assisted support is chain\-based at 6K and direct literal benchmark\-gold presence at 32K after two ambiguous chain labels\. The validated 32K conflict subset has 63/63 chain support at both budgets\.Chain stratification localizes the 6K failures\. No\-conflict accuracy is 100% at BM25\-64 and 88\.5% at BM25\-512, while conflict accuracy falls from 25\.7% to 12\.2%\. Since both the obsolete and latest statements are already retrieved at every 6K budget, this decline is not caused by newly retrieving the competitor\. Additional evidence appears to compound an already difficult selection\. Because conflict status is a property of the question, this comparison is observational\. A separate gold\-assisted diagnostic reaches 97%, but it uses a different prompt and gold\-derived evidence selection\. We treat it only as evidence that short, targeted contexts can be answerable\. Appendix[C](https://arxiv.org/html/2610.36308#A3)reports the preliminary and gold\-assisted diagnostics, Appendix[D](https://arxiv.org/html/2610.36308#A4)gives the complete 6K budget analysis, and Appendix[E](https://arxiv.org/html/2610.36308#A5)gives the full conflict\-stratified analysis\.

### 3\.2An obsolete fact is more harmful than an unrelated fact

The equal\-size intervention gives the central result \(Figure[1](https://arxiv.org/html/2610.36308#S3.F1)\)\. Qwen2\.5 answers all 74 questions correctly in condition A but only 33 in condition B, a 55\.4\-point drop\. The paired comparison has 41 questions correct only in A and none correct only in B \(p=9\.09×10−13p=9\.09\\times 10^\{\-13\}\)\. All 41 harmful flips are classified as stale\. Qwen3 is considerably more robust, but the same direction remains\. Its accuracy falls from 100% in condition A to 87\.8% in condition B, with nine A\-only correct cases, no B\-only correct cases, and nine stale selections \(p=0\.0039p=0\.0039\)\.

![Refer to caption](https://arxiv.org/html/2610.36308v1/plots/figure1_controlled_models.png)Figure 1:Equal\-size intervention on 74 conflict questions\. Replacing one unrelated statement with the obsolete target fact harms both models\.For example, both models answer “Thomas Kyd” for*The Marriage of Figaro*in condition A, but can switch to “Pierre Beaumarchais” once the older statement replaces an unrelated memory, even though the newer statement remains visible and the prompt specifies the chronology\. Appendix[F](https://arxiv.org/html/2610.36308#A6)reports the earlier abstaining\-prompt experiment, while Appendix[G](https://arxiv.org/html/2610.36308#A7)gives the frozen\-prompt pilot, complete cross\-model results, and all paired flips\.

### 3\.332K replication and manual audit

The larger 32K history shows the same overall direction\. The official 100\-question scores are 69% for BM25\-64 and 58% for BM25\-512\. 18 questions are correct only at 64 tokens and seven are correct only at 512 \(p=0\.0433p=0\.0433\)\. All 18 harmful flips are classified as correct\-to\-stale\. Manual inspection identified two ambiguous cases among automatically labeled no\-conflict questions\. We keep the official scores as primary results\. Excluding only those two for sensitivity givesn=98n=98, 69/98 \(70\.4%\) versus 57/98 \(58\.2%\), with 18 versus six discordant successes \(p=0\.0227p=0\.0227\)\.

Manual metric review found both a false negative \(“goalie” for gold “goaltender”\) and a false positive \(“rugby union” contains gold “rugby” despite matching the obsolete value\), plus two final Qwen2\.5 outputs containing both latest and stale strings\. We keep the official substring scores\. Appendices[H](https://arxiv.org/html/2610.36308#A8)and[I](https://arxiv.org/html/2610.36308#A9)give the full 32K replication and the manual audit\.

Table 2:Representative audited outputs\. The benchmark truth is the later synthetic memory, not ordinary world knowledge\.

## 4Discussion and conclusion

#### Answers to the research questions\.

Larger budgets lower accuracy even when the latest statement is present \(RQ1\)\. The loss concentrates on conflict questions \(RQ2\) and at equal size, replacing an unrelated statement with the obsolete competitor harms both models, with every harmful flip selecting the stale value \(RQ3\)\.

#### Evaluation and system implications\.

Retrieving an update does not guarantee that the model will use it\. Evaluation should therefore distinguish whether the latest evidence is retrieved, whether obsolete evidence is also retrieved, and which version the reader selects\. Evidence composition matters beyond token count\. An obsolete target statement is more harmful than an unrelated statement\. These results motivate memory systems that retire, down\-weight, or explicitly mark superseded facts as inactive while preserving provenance and validity information during retrieval\.

#### Model variation and mechanism\.

Qwen3’s smaller drop suggests that model choice may affect robustness, although its nine harmful flips are still stale selections\. In this benchmark, obsolete values often match familiar world knowledge while the synthetic updates contradict it\. Our results therefore establish susceptibility to these competitors, not that temporal obsolescence alone drives the effect\. Counterbalancing chronology or using unfamiliar entities would separate the two\. Our behavioral intervention does not establish whether the cause is attention allocation, prior strength, or decoding dynamics, so we avoid a mechanistic claim\.

#### Scope and broader impact\.

Our study is limited to two open\-weight Qwen models, deterministic decoding, one synthetic benchmark, single\-hop chains of length at most two, and lexical retrieval\. We therefore do not establish the same behavior for dense retrieval, multi\-hop memory, or end\-to\-end systems such as Mem0\. Both models are from the Qwen family\. Cross\-family generalization is untested, and longer update chains or multi\-hop dependencies may compound the effect\. The gold\-assisted audits are diagnostic rather than deployable retrieval evaluations, and the questions share a common benchmark history, which limits population\-level statistical interpretation\. Appendix[J](https://arxiv.org/html/2610.36308#A10)documents the excluded runs and result\-verification checks\. In deployed agents, stale sensitive or operational information could lead to outdated recommendations or disclosures\. Version\-aware retirement can mitigate this risk\. Within these boundaries,when both versions are present, obsolete evidence can override a retrieved correction\.

## Acknowledgments and Disclosure of Funding

We thank the anonymous reviewers of the PALM workshop for their constructive feedback\. Experiments were run on GPU resources provided by Otto\-von\-Guericke\-Universität Magdeburg\.

## References

- Chaoet al\.\(2026\)H\. Chao, Y\. Bai, R\. Sheng, T\. Li, and Y\. SunSTALE: can LLM agents know when their memories are no longer valid?\.CoRRabs/2605\.06527\.External Links:[Link](https://doi.org/10.48550/arXiv.2605.06527),[Document](https://dx.doi.org/10.48550/ARXIV.2605.06527),2605\.06527Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Chhikaraet al\.\(2025\)P\. Chhikara, D\. Khant, S\. Aryan, T\. Singh, and D\. YadavMem0: building production\-ready AI agents with scalable long\-term memory\.InECAI 2025 \- 28th European Conference on Artificial Intelligence, 25\-30 October 2025, Bologna, Italy \- Including 14th Conference on Prestigious Applications of Intelligent Systems \(PAIS 2025\),I\. Lynce, N\. Murano, M\. Vallati, S\. Villata, F\. Chesani, M\. Milano, A\. Omicini, and M\. Dastani \(Eds\.\),Frontiers in Artificial Intelligence and Applications, Vol\.413,pp\. 2993–3000\.External Links:[Link](https://doi.org/10.3233/FAIA251160),[Document](https://dx.doi.org/10.3233/FAIA251160)Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Huet al\.\(2026\)Y\. Hu, Y\. Wang, and J\. J\. McAuleyEvaluating memory in LLM agents via incremental multi\-turn interactions\.InThe Fourteenth International Conference on Learning Representations, ICLR 2026,External Links:[Link](https://iclr.cc/virtual/2026/poster/10010781)Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2610.36308#S1.p2.1)\.
- Kurfali and Östling \(2025\)M\. Kurfali and R\. ÖstlingConflicting needles in a haystack: how LLMs behave when faced with contradictory information\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 34361–34376\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.1742/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.1742),ISBN 979\-8\-89176\-332\-6Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Lewiset al\.\(2020\)P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. KielaRetrieval\-augmented generation for knowledge\-intensive NLP tasks\.InAdvances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6\-12, 2020, virtual,H\. Larochelle, M\. Ranzato, R\. Hadsell, M\. Balcan, and H\. Lin \(Eds\.\),External Links:[Link](https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html)Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Liuet al\.\(2024\)N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. LiangLost in the middle: how language models use long contexts\.Trans\. Assoc\. Comput\. Linguistics12,pp\. 157–173\.External Links:[Link](https://doi.org/10.1162/tacl/_a/_00638),[Document](https://dx.doi.org/10.1162/TACL%5FA%5F00638)Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Maharanaet al\.\(2024\)A\. Maharana, D\. Lee, S\. Tulyakov, M\. Bansal, F\. Barbieri, and Y\. FangEvaluating very long\-term conversational memory of LLM agents\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\), ACL 2024, Bangkok, Thailand, August 11\-16, 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),pp\. 13851–13870\.External Links:[Link](https://doi.org/10.18653/v1/2024.acl-long.747),[Document](https://dx.doi.org/10.18653/V1/2024.ACL-LONG.747)Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Reddy and Challaram \(2026\)V\. Reddy and S\. ChallaramDon’t ask the LLM to track freshness: A deterministic recipe for memory conflict resolution\.CoRRabs/2606\.01435\.External Links:[Link](https://doi.org/10.48550/arXiv.2606.01435),[Document](https://dx.doi.org/10.48550/ARXIV.2606.01435),2606\.01435Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Robertson and Zaragoza \(2009\)S\. E\. Robertson and H\. ZaragozaThe probabilistic relevance framework: BM25 and beyond\.Found\. Trends Inf\. Retr\.3\(4\),pp\. 333–389\.External Links:[Link](https://doi.org/10.1561/1500000019),[Document](https://dx.doi.org/10.1561/1500000019)Cited by:[§2](https://arxiv.org/html/2610.36308#S2.SS0.SSS0.Px2.p1.1)\.
- Shiet al\.\(2023\)F\. Shi, X\. Chen, K\. Misra, N\. Scales, D\. Dohan, E\. H\. Chi, N\. Schärli, and D\. ZhouLarge language models can be easily distracted by irrelevant context\.InInternational Conference on Machine Learning, ICML 2023, 23\-29 July 2023, Honolulu, Hawaii, USA,A\. Krause, E\. Brunskill, K\. Cho, B\. Engelhardt, S\. Sabato, and J\. Scarlett \(Eds\.\),Proceedings of Machine Learning Research, Vol\.202,pp\. 31210–31227\.External Links:[Link](https://proceedings.mlr.press/v202/shi23a.html)Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Team \(2025\)Q\. TeamQwen3 technical report\.CoRRabs/2505\.09388\.External Links:[Link](https://doi.org/10.48550/arXiv.2505.09388),[Document](https://dx.doi.org/10.48550/ARXIV.2505.09388),2505\.09388Cited by:[§2](https://arxiv.org/html/2610.36308#S2.SS0.SSS0.Px4.p1.1)\.
- Wallatet al\.\(2026\)J\. Wallat, W\. Nejdl, and S\. SikdarWhen facts change: temporal knowledge conflict resolution in llms\.InFindings of the Association for Computational Linguistics, ACL 2026, San Diego, California, United States, July 2\-7, 2026,M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),pp\. 2154–2184\.External Links:[Link](https://doi.org/10.18653/v1/2026.findings-acl.103),[Document](https://dx.doi.org/10.18653/V1/2026.FINDINGS-ACL.103)Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Wuet al\.\(2025\)D\. Wu, H\. Wang, W\. Yu, Y\. Zhang, K\. Chang, and D\. YuLongMemEval: benchmarking chat assistants on long\-term interactive memory\.InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24\-28, 2025,External Links:[Link](https://openreview.net/forum?id=pZiyCaVuti)Cited by:[§1](https://arxiv.org/html/2610.36308#S1.SS0.SSS0.Px1.p1.1)\.
- Yanget al\.\(2024\)A\. Yang, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Li, D\. Liu, F\. Huang, H\. Wei, H\. Lin, J\. Yang, J\. Tu, J\. Zhang, J\. Yang, J\. Yang, J\. Zhou, J\. Lin, K\. Dang, K\. Lu, K\. Bao, K\. Yang, L\. Yu, M\. Li, M\. Xue, P\. Zhang, Q\. Zhu, R\. Men, R\. Lin, T\. Li, T\. Xia, X\. Ren, X\. Ren, Y\. Fan, Y\. Su, Y\. Zhang, Y\. Wan, Y\. Liu, Z\. Cui, Z\. Zhang, and Z\. QiuQwen2\.5 technical report\.CoRRabs/2412\.15115\.External Links:[Link](https://doi.org/10.48550/arXiv.2412.15115),[Document](https://dx.doi.org/10.48550/ARXIV.2412.15115),2412\.15115Cited by:[§2](https://arxiv.org/html/2610.36308#S2.SS0.SSS0.Px4.p1.1)\.

## Supplementary material

## Appendix AReproducibility and evaluation protocol

### A\.1Dataset and chronological semantics

All experiments use theConflict\_Resolutionsplit of MemoryAgentBench\. Our main source usesfactconsolidation\_sh\_6k, which contains 455 numbered statements and 100 questions\. The robustness source isfactconsolidation\_sh\_32k, with 2,310 statements and 100 questions\. Statement numbers define the chronology: when two statements conflict about the same relation, the higher\-numbered statement is treated as the newer version\. Because the benchmark facts are synthetic, ordinary world knowledge is not used as the evaluation target\. The MemoryAgentBench code repository is released under the MIT License\.

### A\.2Models, Ollama, and decoding

We evaluateqwen2\.5:7b\-instructandqwen3:14b; the corresponding public Qwen checkpoints are Apache\-2\.0 licensed\. They were served by Ollama on a remote university GPU system, and our local Python scripts sent requests through the Ollama HTTP chat API over a VPN\. The released code omits private hostnames, usernames, and credentials\. The full\-history 6K smoke test and 100\-question full\-context baseline usenum\_ctx=16384\. The question\-only baseline, three\-question oracle diagnostic, and 6K BM25 retrieval runs usenum\_ctx=4096\. The final controlled runs and the corrected 32K BM25 replication use temperature 0, seed 42,num\_ctx=2048, andnum\_predict=64\. Qwen3 additionally sets the top\-level API optionthink=false\.

### A\.3Frozen final prompt

> You are answering a synthetic memory benchmark\. Use only the supplied numbered memory statements as the source of truth, even when they contradict ordinary world knowledge\. The statements are chronological: higher statement numbers are newer and replace lower conflicting statements about the same fact\. Return only the shortest possible answer phrase\. Do not explain\.

### A\.4Retrieval and support audits

BM25 ranks individual statements from the complete history\. We add statements in ranked order until adding another statement would exceed the evidence\-token budget, then restore the selected statements to chronological order before presenting them to the model\. The 6K experiments use budgets of 64, 128, 256, and 512 tokens while the 32K replication uses 64 and 512\. After generation, gold\-assisted audits uses the benchmark answers to identify the intended latest statement\. They diagnose support presence but are not deployable retrieval metrics\. At 32K, manual inspection showed that the automatic chain template mislabeled two questions\. Direct inspection finds the benchmark\-gold statement in all 100 retrieved contexts at both budgets, and the validated 63\-question conflict subset has chain\-validated latest support in 63/63 contexts at both budgets\.

### A\.5Scoring and paired tests

The official outcome is normalized substring exact match\. Predictions are also classified as latest, stale, unknown, or other by comparison with reconstructed update\-chain values\. We use exact two\-sided McNemar tests for paired comparisons of the same benchmark questions under different conditions\. Because questions share a single history, thesepp\-values describe this benchmark instance rather than a population of independent histories\. Appendix[I](https://arxiv.org/html/2610.36308#A9)separately documents identified semantic false positives/negatives and mixed\-string outputs without silently altering the benchmark metric\.

## Appendix BComplete experiment inventory

Table 3:Complete experiment inventory\. Numbering follows the saved result files\.IDPurposeSetupSizeMain observationStatus02Full\-context smokeQwen2\.5; full 6K history5 questions2/5 \(40%\)Preliminary03First\-five evidence auditInspect full\-history outputs5 questionsGold/latest text present in failed casesDiagnostic04Oracle prompt diagnosticQuestion\-only / latest\-only / all relevant3 x 30/3; 3/3; 3/3Diagnostic05Full\-context baselineQwen2\.5; full 6K history100 calls37/100Valid baseline06Question\-only baselineQwen2\.5; no supplied memory100 calls19/100Valid baseline; prompt differs07Full vs\. question\-only pairingPaired 05 and 06100 pairs19 both; 18 full\-only; 0 question\-only\-onlyDiagnostic08\-09BM25\-512 \+ support auditQwen2\.5; 6K100 calls32/100; gold\-assisted support 100/100Valid \+ diagnostic audit10Gold\-assisted isolated contextQwen2\.5; gold\-derived one\-statement evidence100 calls97/100Diagnostic only; different prompt11\-12BM25 budget sweep64/128/256/512 tokens400 calls45/40/36/32%Valid13Budget trajectories and pairsPaired BM25 conditions100 trajectories64 vs 512: 14 vs 1; p=0\.00098Valid diagnostic14Update\-chain reconstruction6K fact templates100 questions100 resolved; 74 conflictGold\-assisted diagnostic15Conflict stratificationAll baseline/retrieval conditions100 questionsFailures concentrated in conflict casesDiagnostic16\-17Exploratory conflict interventionEarlier abstaining prompt348 calls65% unrelated vs 12\.2% old\+latest\+unrelatedExploratory; prompt\-sensitive18Frozen\-prompt pilotFirst 10 conflict questions40 calls10/10 vs 3/10; p=0\.0156Pilot19Final Qwen2\.5 interventionFrozen equal\-size conditions148 calls74/74 vs 33/74; 41 stale flipsFinal confirmatory20bFinal Qwen3 interventionFrozen equal\-size; think=false148 calls74/74 vs 65/74; 9 stale flipsFinal robustness21b32K BM25 replicationQwen2\.5; budgets 64 and 512200 calls69/100 vs 58/100; 18 harmful flipsFinal robustness22Manual auditCross\-run qualitative review20 cases13 sampled flips genuine; 1 metric miss; 2 ambiguitiesFinal auditThe experimental sequence were designed as diagnostic steps\. preliminary baselines suggested that memory helped but it remained limited\. Increasing the retrieval budget showed a degradation trend\. chain reconstruction localized the trend to conflict questions\. The equal\-size intervention isolated the effect of the obsolete competing fact\. Qwen3 and 32K runs tested the robustness and the manual audit checked semantic and chain\-label edge cases\.

## Appendix CPreliminary baselines and oracle diagnostics

Table 4:Preliminary and baseline results\.The full 6K history improves over question only prompting by 18 percentage points\. In the paired comparison, 19 questions are answered correctly under both conditions, 18 correct only with the full history, zero correct only without memory, and 63 wrong under both\. This comparison provides context but should not be interpreted as a causal effect of memory\.

Table 5:First\-five full\-context evidence audit\. “Gold shown" counts literal occurrences displayed by the audit, not relation\-specific support\.The first\-five inspection provided the earliest sign that the model could answer incorrectly even when the updated value existed in the supplied history\. This finding led us to reconstruct relation\-specific update chains instead of relying only on whether the answer string appeared somewhere in the history\.

Table 6:Three\-question oracle diagnostic\.In the three\-question pilot, the model can answer all the three cases correctly when supplied with the isolated latest statement or the complete relevant chain in chronological order\. The later 100\-question gold\-assisted diagnostic run reaches 97%, but it used the gold answer to select the evidence and a different prompt from the final frozen experiments\. We therefore do not compare its absolute accuracy directly with final controlled conditions\.

## Appendix DComplete 6K evidence\-budget analysis

Table 7:Complete 6K BM25 budget sweep\. Support is the gold\-assisted latest\-statement diagnostic\.![Refer to caption](https://arxiv.org/html/2610.36308v1/plots/figure2_budget_accuracy.png)Figure 2:Accuracy as a function of BM25 evidence budget\. The 6K source provides the full fourbudget sweep, 32K repeats the endpoints\.Table 8:All paired 6K BM25 budget comparisons\.Table 9:Per\-question correctness trajectories across budgets 64, 128, 256, and 512\. A one denotes correct\.The monotone aggregate curve does not mean that every individual question worsens\. 29 questions are correct at all four budgets, 49 are wrong at all four, and the remaining 22 move between correct and incorrect states\. The most informative endpoint comparison has 14 losses and one gain\.

Table 10:All discordant 6K BM25\-64 versus BM25\-512 questions\.QQuestionGold/latestAnswer 64Answer 512Direction1Which sport is goaltender associated with?pesäpallopesäpalloice hockey64\-only14What is the country of citizenship of Cédric Klapisch?MalaysiaMalaysiaFrance64\-only18Which country was association football created in?ItalyItalyEngland64\-only24Who is the director of American Broadcasting Company?Ernst HeinkelErnst HeinkelBob Iger64\-only25Which city did Gennady Rozhdestvensky die in?United States of AmericaUnited States of AmericaMoscow64\-only29What type of music does Michael Mantler play?post\-punkpost\-punkjazz64\-only54Which religion is Imelda Marcos affiliated with?atheismatheismCatholicism64\-only55What type of music does Dana International play?Australian hip hopAustralian hip hopUNKNOWN64\-only60Which sport is Bo Ryan associated with?association footballbasketballassociation football512\-only61What type of music does John McVie play?rap rockrap rockrock music64\-only73What position does Blair Walsh play?placekickerplacekickerUNKNOWN64\-only84Which religion is Karen Armstrong affiliated with?Church of ScotlandChurch of ScotlandCatholic Church64\-only94What is the country of citizenship of Charles Messier?IranIranFrance64\-only96What position does Mike D’Antoni play?point guardpoint guardUNKNOWN64\-only100What position does Irving Fryar play?wide receiverwide receiverUNKNOWN64\-only
## Appendix EUpdate chains and conflict localization

The 6K chain audit successfully reconstructs all 100 questions\. Of these, 74 have one older conflicting value, while 26 have no previous conflict, and every chain contains either one or two statements\. The update\-chain relation templates and the latest statement are identified with gold\-assisted post\-hoc matching, this analysis is diagnostic rather than deployable\.

Table 11:Complete 6K conflict stratification\.Table 12:Paired BM25\-64 versus BM25\-512 by chain stratum\.The question\-only model receives no benchmark memory\. It answers none of the 74 synthetic conflict update questions correctly, while it answers 19 of the 26 no\-conflict questions correctly, which is consistent with reliance on pretrained knowledge\. As the retrieval budget grows, most of the decline occurs in the conflict stratum\. Between 64 and 512 tokens, conflict questions show a net loss of 10 correct answers, compared with a net loss of three among no\-conflict questions\. However, both the latest and old target statements are already present in all 74 conflict contexts at every budget\. Therefore, Larger budgets therefore do not newly introduce the obsolete competitor, instead, the results suggest that competing versions make the model more weak when additional evidence is added\. The later equal\-size intervention tests the effect of the obsolete competitor directly\.

## Appendix FExploratory abstaining\-prompt experiment

Before freezing the final prompt, we ran an exploratory four\-condition experiment using an earlier prompt that encouragedUNKNOWN\. These outputs are useful for understanding prompt sensitivity but we do not treat them as our principal evidence because the high abstention rate obscures stale selection\.

Table 13:Exploratory controlled experiment under the earlier abstaining prompt\.Table 14:Answer classifications for the 74 conflict questions under the earlier abstaining prompt\.The exploratory equal\-size comparison, the model received either the latest statement plus three unrelated statements or the old and latest statements plus two unrelated statements yields 31 first\-only correct and one second\-only correct \(p=1\.54×10−8p=1\.54\\times 10^\{\-8\}\)\. However, the model returnedUNKNOWNfor 34 questions in the first condition and 38 in the second\. The response “goalie” for gold “goaltender” is categorized as other by substring matching but is a semantic synonym\. These findings led us to adopt the final non\-abstaining frozen prompt\.

## Appendix GFrozen prompt pilot and final cross\-model intervention

Table 15:Frozen\-prompt pilot on the first 10 conflict questions\.In the pilot’s equal\-size comparison has seven latest\-plus\-unrelated\-only successes and no successes in the opposite direction \(exactp=0\.015625p=0\.015625\)\. We used this pilot only to check that the non\-abstaining prompt produced short, usable answers before the full run\.

Table 16:Final frozen equal\-size intervention\. Condition A is latest plus three unrelated statements; B is old plus latest plus two unrelated statements\.The final Qwen2\.5 and Qwen3 experiments each contain 148 completed API calls and no outputs classified as unknown or other\. Within each model, mean prompt\-evaluation counts differ by less than one token between A and B, confirming approximately equal prompt size\. Under the official classifier, all 50 harmful paired flips are stale selections\. Appendix[I](https://arxiv.org/html/2610.36308#A9)records two Qwen2\.5 mixed\-string outputs that are not represented by the four\-way label scheme and reports a strict sensitivity analysis\.

Table 17:All 50 correct\-to\-stale flips in the final equal\-size interventions\.ModelQQuestionLatestObsoleteAnswer AAnswer BQwen2\.5\-7B2What is Nobuhiro Watsuki famous for?The Fairly OddParentsRurouni KenshinThe Fairly OddParentsRurouni KenshinQwen2\.5\-7B6Who is the author of The Marriage of Figaro?Thomas KydPierre BeaumarchaisThomas KydPierre BeaumarchaisQwen2\.5\-7B7Which sport is Tunisia national football team associated with?basketballassociation footballbasketballassociation footballQwen2\.5\-7B8Which sport is quarterback associated with?Muay ThaiAmerican footballMuay ThaiAmerican footballQwen2\.5\-7B9What is the official language of Japan?SwedishJapaneseSwedishJapaneseQwen2\.5\-7B13What position does Robert Parish play?quarterbackcenterquarterbackcenterQwen2\.5\-7B14What is the country of citizenship of Cédric Klapisch?MalaysiaFranceMalaysiaFranceQwen2\.5\-7B15Who is the author of Sidereus Nuncius?Samuel BeckettGalileo GalileiSamuel BeckettGalileo GalileiQwen2\.5\-7B16Which continent is Germany located in?AfricaEuropeAfricaEuropeQwen2\.5\-7B17Which religion is Kingdom of Ireland affiliated with?ZoroastrianismCatholic ChurchZoroastrianismCatholic ChurchQwen2\.5\-7B20Who is the chief executive officer of Microsoft?Steve JobsSatya NadellaSteve JobsSatya NadellaQwen2\.5\-7B30Which city did Oscar Wilde die in?GuangzhouParisGuangzhouParisQwen2\.5\-7B31Which country was baseball created in?JapanUnited States of AmericaJapanUnited States of AmericaQwen2\.5\-7B34Which country was rap rock created in?Russian EmpireUnited States of AmericaRussian EmpireUnited States of AmericaQwen2\.5\-7B35Where was Christianity founded?TaipeiJerusalemTaipeiJerusalemQwen2\.5\-7B36What is the official language of United States of America?GermanAmerican EnglishGermanAmerican EnglishQwen2\.5\-7B39What is the capital of France?HarareParisHarare \(from statement 171, replaced by a false statement\) ParisParisQwen2\.5\-7B40Who is the developer of SteamOS?ABB GroupValve CorporationABB GroupValve CorporationQwen2\.5\-7B41Which city is the headquarter of University of Bucharest located in?AnkaraBucharestAnkaraBucharestQwen2\.5\-7B44What is the capital of Australia?OderzoCanberraOderzoCanberraQwen2\.5\-7B45Which continent is Malaysia located in?AntarcticaAsiaAntarcticaAsiaQwen2\.5\-7B49Which country was basketball created in?Soviet UnionUnited States of AmericaSoviet UnionUnited States of AmericaQwen2\.5\-7B52Which sport is Slovakia national football team associated with?cricketassociation footballcricketassociation footballQwen2\.5\-7B57Which country was Charlie Hebdo created in?ItalyFranceItalyFranceQwen2\.5\-7B58Who is Elvis Presley married to?Charles the BoldPriscilla PresleyCharles the BoldPriscilla PresleyQwen2\.5\-7B59Who performed Hermione Granger?Kylie MinogueEmma WatsonKylie MinogueEmma WatsonQwen2\.5\-7B60Which sport is Bo Ryan associated with?association footballbasketballassociation footballbasketballQwen2\.5\-7B62What position does Rogério Ceni play?flankergoalkeeperflankergoalkeeperQwen2\.5\-7B64What language does Kylie Minogue speak?GermanEnglishGermanEnglishQwen2\.5\-7B66Which sport is placekicker associated with?rugbyAmerican footballrugbyAmerican footballQwen2\.5\-7B74Who is Tony Blair married to?Alan Jay LernerCherie BlairAlan Jay LernerCherie BlairQwen2\.5\-7B78Who is the chairperson of Congolese Party of Labour?Stephen McNeilDenis Sassou\-NguessoStephen McNeilDenis Sassou\-NguessoQwen2\.5\-7B81What position does Lisa Leslie play?goaltendercentergoaltendercenterQwen2\.5\-7B82What is the name of the current head of state in Soviet Union?Elizabeth IIMikhail GorbachevElizabeth IIMikhail GorbachevQwen2\.5\-7B86What is the capital of Romania?RajanpurBucharestRajanpurBucharestQwen2\.5\-7B87Which continent is India located in?AfricaAsiaAfricaAsiaQwen2\.5\-7B88Which city did Olga of Kiev die in?RodezKyivRodezKyivQwen2\.5\-7B89Who founded Sasanian Empire?Walter ChryslerArdashir IWalter ChryslerArdashir IQwen2\.5\-7B91Which sport is Surrey Senior Cup associated with?basketballassociation footballbasketballassociation footballQwen2\.5\-7B92What is the country of citizenship of Narendra Modi?AustraliaIndiaAustraliaIndiaQwen2\.5\-7B99Who is Charles Darwin married to?Amala PaulEmma DarwinAmala PaulEmma DarwinQwen3\-14B6Who is the author of The Marriage of Figaro?Thomas KydPierre BeaumarchaisThomas Kyd\.Pierre Beaumarchais\.Qwen3\-14B9What is the official language of Japan?SwedishJapaneseSwedish\.Japanese\.Qwen3\-14B44What is the capital of Australia?OderzoCanberraOderzo\.Canberra\.Qwen3\-14B57Which country was Charlie Hebdo created in?ItalyFranceItaly\.France\.Qwen3\-14B58Who is Elvis Presley married to?Charles the BoldPriscilla PresleyCharles the Bold\.Priscilla Presley\.Qwen3\-14B59Who performed Hermione Granger?Kylie MinogueEmma WatsonKylie Minogue\.Emma Watson\.Qwen3\-14B71Which company is AP1000 produced by?CanadairWestinghouse ElectricCanadair\.Westinghouse Electric\.Qwen3\-14B89Who founded Sasanian Empire?Walter ChryslerArdashir IWalter Chrysler\.Ardashir I\.Qwen3\-14B92What is the country of citizenship of Narendra Modi?AustraliaIndiaAustralia\.India\.
## Appendix HComplete 32K replication

The valid 32K experiment contains 200 completed calls\. 100 questions at BM25\-64 and the same 100 questions with BM25\-512\. The automated chain audit initially labels 63 conflict and 37 no\-conflict questions\. Manual review found two questions where the gold\-derived chain template attached to an unrelated sentence\. We flag these cases as ambiguous rather than silently changing their labels\.

Table 18:Complete 32K BM25 replication\. Automated chain support is shown here; direct literal support is 100/100 for both budgets after manual audit\.The official aggregate accuracies are 69/100 \(69%\) for BM25\-64 and 58/100 \(58%\) for BM25\-512 \. Direct inspection finds the statement supporting the benchmark\-gold statement in 100/100 contexts under each budget\. After excluding the two ambiguous cases, the clean no\-conflict subset contains 35 questions, with accuracies of 34/35 \(97\.1%\) for BM25\-64 and 35/35 \(100%\) for BM25\-512\. Excluding the same two cases only for an overall sensitivity analysis givesn=98n=98: 69/98 \(70\.4%\) versus 57/98 \(58\.2%\), with 18 BM25\-64\-only correct, six BM25\-512\-only correct, 51 both correct, and 23 both wrong \(exactp=0\.02266p=0\.02266\)\. These sensitivity values do not replace the official 100\-question scores\.

Table 19:All 18 detrimental BM25\-64 to BM25\-512 flips in the 32K replication\.QQuestionLatestObsoleteAnswer 64Answer 512Type2What position does Nick Rimando play?flankergoalkeeperflankergoalkeeperstale12What is the country of citizenship of Colin Irwin?IsraelUnited KingdomIsraelUnited Kingdomstale20Which religion is Maria Amalia of Naples and Sicily affiliated with?Church of GreeceCatholic ChurchChurch of GreeceCatholic Churchstale21What position does Tony Parker play?shooting guardpoint guardshooting guardpoint guardstale42Who is the chief executive officer of Apple Inc\.?Vijay MallyaTim CookVijay MallyaTim Cookstale48Which continent is San Francisco located in?AfricaNorth AmericaAfricaNorth Americastale54Which sport is Imran Khan associated with?association footballcricketassociation footballcricketstale57Who founded Transcendental Meditation movement?Adalbert of PragueMaharishi Mahesh YogiAdalbert of PragueMaharishi Mahesh Yogistale58What is the name of the current head of state in United States of America?ConnachtaDonald TrumpConnachta\.Donald Trumpstale66Which religion is Bahawalpur State affiliated with?Catholic ChurchIslamCatholic ChurchIslamstale67Who is the chief executive officer of Advanced Micro Devices?Sundar PichaiLisa SuSundar PichaiLisa Sustale69Who is the developer of Internet Explorer 5?GoogleMicrosoftGoogleMicrosoftstale71Who is the author of The Age of Reason?Allen GinsbergThomas PaineAllen GinsbergThomas Painestale78What is the country of citizenship of Cédric Klapisch?MalaysiaFranceMalaysiaFrancestale80Which sport is Sports Interactive associated with?basketballassociation footballbasketballassociation footballstale89Who was Finnish created by?Stirling SilliphantaMikael AgricolaStirling SilliphantaMikael Agricolastale96What is the country of citizenship of Henri Grégoire?IndiaFranceIndiaFrancestale99What is the name of the current head of the South Korea government?Mauro CarlesseMoon Jae\-inMauro CarlesseMoon Jae\-instaleA separate metric audit found one false positive in the 32K results\. For the question “Which sport is flanker associated with?”, the gold answer is “rugby”, the obsolete value is “rugby union”, and the model answers “rugby union” under both budgets\. Because rugby” is a substring of rugby union”, normalized substring matching counts both responses as correct\. Reclassifying this identified case as stale would change the marginal accuracies to 68/100 and 57/100\. It would not change the 64\-versus\-512 paired discordance because the case has the same status under both budgets\.

We do not report a full\-history 32K baseline because the evidence alone tokenizes to approximately 37\.5K tokens\. This already exceeds the server model’s reported 32,768\-token context window before accounting for the system prompt, question, and output allowance\.

## Appendix IManual audit

We selected twenty cases for the original qualitative audit to cover five Qwen2\.5 controlled flips, four Qwen3 controlled flips, four 32K budget flips, two stable correct answers, two exploratory unknown answers, one semantic\-scoring case, and two 32K chain ambiguities\. The audit is qualitative and does not replace the official substring metric\.

Table 20:Complete 20\-case manual audit\.IDCategoryModelQLatestObsoleteAnswer AAnswer B1Qwen 2\.5 correct\-to\-stale flipQwen2\.52The Fairly OddParentsRurouni KenshinThe Fairly OddParentsRurouni Kenshin2Qwen 2\.5 correct\-to\-stale flipQwen2\.56Thomas KydPierre BeaumarchaisThomas KydPierre Beaumarchais3Qwen 2\.5 correct\-to\-stale flipQwen2\.59SwedishJapaneseSwedishJapanese4Qwen 2\.5 correct\-to\-stale flipQwen2\.520Steve JobsSatya NadellaSteve JobsSatya Nadella5Qwen 2\.5 correct\-to\-stale flipQwen2\.535TaipeiJerusalemTaipeiJerusalem6Qwen 3 correct\-to\-stale flipQwen344OderzoCanberraOderzo\.Canberra\.7Qwen 3 correct\-to\-stale flipQwen357ItalyFranceItaly\.France\.8Qwen 3 correct\-to\-stale flipQwen359Kylie MinogueEmma WatsonKylie Minogue\.Emma Watson\.9Qwen 3 correct\-to\-stale flipQwen392AustraliaIndiaAustralia\.India\.1032K BM25 correct\-to\-stale flipQwen2\.52flankergoalkeeperflankergoalkeeper1132K BM25 correct\-to\-stale flipQwen2\.512IsraelUnited KingdomIsraelUnited Kingdom1232K BM25 correct\-to\-stale flipQwen2\.542Vijay MallyaTim CookVijay MallyaTim Cook1332K BM25 correct\-to\-stale flipQwen2\.599Mauro CarlesseMoon Jae\-inMauro CarlesseMoon Jae\-in14Stable correct latest answerQwen2\.51pesäpalloice hockeypesäpallopesäpallo15Stable correct latest answerQwen33IndiaEnglandIndia\.India\.16Exploratory UNKNOWN / prompt sensitivityQwen2\.52The Fairly OddParentsRurouni KenshinUNKNOWN17Exploratory UNKNOWN / prompt sensitivityQwen2\.59SwedishJapaneseUNKNOWN18Semantic scoring artifactQwen2\.581goaltendercentergoalie1932K chain/benchmark ambiguityQwen2\.59LondonAutomated chain label was emptyWashington, D\.C\.Washington, D\.C\.2032K chain/benchmark ambiguityQwen2\.510RomeAutomated chain label was emptyWatertownRomeAll 13 sampled principal flips involved direct selection of an explicitly stated obsolete value\. The final frozen runs contain no answers classified as unknown\. The goalie” response for the gold answer goaltender” is a metric false negative\. The two 32K ambiguities led us to report direct literal support and to exclude from the clean no\-conflict subgroup, but they do not alter the official 100\-question aggregate scores\.

### I\.1Additional metric and mixed\-output checks

A subsequent line\-by\-line review of the raw outputs identified two further scoring caveats\. First, the 32K “flanker” answer “rugby union” is semantically the reconstructed obsolete value but contains the gold string “rugby\.” The official substring metric therefore counts it as correct under both budgets\. Second, two final Qwen2\.5 responses contain both the latest and stale strings\. In Condition A, the response to the France\-capital question contains Harare …Paris,” while in Condition B, the response to the*Our Mutual Friend*question is “Charles Darwin \(incorrect, replaced by Charles Dickens\)\.” Because the gold string appears in each response, the official metric and classifier mark both as latest\.

For transparency, a strict sensitivity analysis that treats these two mixed Qwen2\.5 responses as non\-latest yields 73/74 \(98\.6%\) in A and 32/74 \(43\.2%\) in B\. The paired comparison remains 41 A\-only versus none B\-only, so the exact McNemarpp\-value is unchanged\. Under this stricter analysis, one of the 41 discordant Condition B responses is mixed rather than a pure stale answer\. We retain the official benchmark scores in all primary tables\.

## Appendix JValidity and excluded runs

### J\.1Excluded and diagnostic runs

We exclude the initial Qwen3 experiment because thinking mode used the 64\-token generation allowance before producing a usable final answer\. The valid Qwen3 experiment therefore setsthink=falseand is reported separately in the20bresult files\.

We also exclude the attempted full\-history 32K condition\. That run used an 8,192\-token context setting, while the evidence alone tokenized to approximately 37\.5K tokens\. The evidence therefore exceeded both the selected context window and the server model’s reported maximum context limit of 32,768 tokens, even before accounting for the instructions, question, and output allowance\. The reported 32K results therefore use only the corrected BM25\-64 and BM25\-512 runs from21b\.

The gold\-assisted isolated and support experiments are used only as diagnostics because the benchmark answers are used after generation either to select evidence or to identify the intended supporting evidence\. The earlier abstaining\-prompt experiment is likewise reported only as a prompt\-sensitivity analysis and is not combined with the final non\-abstaining results\.

### J\.2Result verification

For the canonical runs, we cross\-checked the archived raw outputs against their summary files for the reported call counts, accuracies, response classifications, token statistics, paired comparisons, and exact McNemar values\. We checked the archived summaries of the excluded runs separately against their recorded configurations and exclusion rationale\. We found no inconsistencies in these checks\. Private server addresses and credentials are not included in the archive\.

## Appendix KNeurIPS paper checklist

## NeurIPS Paper Checklist

1. 1\.Claims
2. Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope?
3. Answer:\[Yes\]
4. Justification: The abstract and Introduction state the benchmark setting, controlled intervention, empirical findings, and limits of generalization\.
5. Guidelines: - •The answer\[N/A\]means that the abstract and introduction do not include the claims made in the paper\. - •The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations\. A\[No\]or\[N/A\]answer to this question will not be perceived well by the reviewers\. - •The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings\. - •It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper\.
6. 2\.Limitations
7. Question: Does the paper discuss the limitations of the work performed by the authors?
8. Answer:\[Yes\]
9. Justification: The Discussion and conclusion section discusses dataset, model, retrieval, metric, statistical, and compute limitations\.
10. Guidelines: - •The answer\[N/A\]means that the paper has no limitation while the answer\[No\]means that the paper has limitations, but those are not discussed in the paper\. - •The authors are encouraged to create a separate “Limitations” section in their paper\. - •The paper should point out any strong assumptions and how robust the results are to violations of these assumptions \(e\.g\., independence assumptions, noiseless settings, model well\-specification, asymptotic approximations only holding locally\)\. The authors should reflect on how these assumptions might be violated in practice and what the implications would be\. - •The authors should reflect on the scope of the claims made, e\.g\., if the approach was only tested on a few datasets or with a few runs\. In general, empirical results often depend on implicit assumptions, which should be articulated\. - •The authors should reflect on the factors that influence the performance of the approach\. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting\. Or a speech\-to\-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon\. - •The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size\. - •If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness\. - •While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper\. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community\. Reviewers will be specifically instructed to not penalize honesty concerning limitations\.
11. 3\.Theory assumptions and proofs
12. Question: For each theoretical result, does the paper provide the full set of assumptions and a complete \(and correct\) proof?
13. Answer:\[N/A\]
14. Justification: The paper contains no theoretical theorems or proofs\.
15. Guidelines: - •The answer\[N/A\]means that the paper does not include theoretical results\. - •All the theorems, formulas, and proofs in the paper should be numbered and cross\-referenced\. - •All assumptions should be clearly stated or referenced in the statement of any theorems\. - •The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition\. - •Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material\. - •Theorems and Lemmas that the proof relies upon should be properly referenced\.
16. 4\.Experimental result reproducibility
17. Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper \(regardless of whether the code and data are provided or not\)?
18. Answer:\[Yes\]
19. Justification: Section 2 and Appendix A disclose the dataset sources, prompts, model tags, stage\-specific context settings, final inference parameters, retrieval procedure, metrics, and exclusions\.
20. Guidelines: - •The answer\[N/A\]means that the paper does not include experiments\. - •If the paper includes experiments, a\[No\]answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not\. - •If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable\. - •Depending on the contribution, reproducibility can be accomplished in various ways\. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model\. In general\. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model \(e\.g\., in the case of a large language model\), releasing of a model checkpoint, or other means that are appropriate to the research performed\. - •While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution\. For example 1. \(a\)If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm\. 2. \(b\)If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully\. 3. \(c\)If the contribution is a new model \(e\.g\., a large language model\), then there should either be a way to access this model for reproducing the results or a way to reproduce the model \(e\.g\., with an open\-source dataset or instructions for how to construct the dataset\)\. 4. \(d\)We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility\. In the case of closed\-source models, it may be that access to the model is limited in some way \(e\.g\., to registered users\), but it should be possible for other researchers to have some path to reproducing or verifying the results\.
21. 5\.Open access to data and code
22. Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
23. Answer:\[Yes\]
25. Guidelines: - •The answer\[N/A\]means that paper does not include experiments requiring code\. - • - •While we encourage the release of code and data, we understand that this might not be possible, so\[No\]is an acceptable answer\. Papers cannot be rejected simply for not including code, unless this is central to the contribution \(e\.g\., for a new open\-source benchmark\)\. - •The instructions should contain the exact command and environment needed to run to reproduce the results\. See the NeurIPS code and data submission guidelines \([https://neurips\.cc/public/guides/CodeSubmissionPolicy](https://neurips.cc/public/guides/CodeSubmissionPolicy)\) for more details\. - •The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc\. - •The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines\. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why\. - •At submission time, to preserve anonymity, the authors should release anonymized versions \(if applicable\)\. - •Providing as much information as possible in supplemental material \(appended to the paper\) is recommended, but including URLs to data and code is permitted\.
26. 6\.Experimental setting/details
27. Question: Does the paper specify all the training and test details \(e\.g\., data splits, hyperparameters, how they were chosen, type of optimizer\) necessary to understand the results?
28. Answer:\[Yes\]
29. Justification: Section 2 reports the split, statement and question counts, evidence construction, final model settings, prompt, and evaluation procedure; Appendix A gives stage\-specific context settings\.
30. Guidelines: - •The answer\[N/A\]means that the paper does not include experiments\. - •The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them\. - •The full details can be provided either with the code, in appendix, or as supplemental material\.
31. 7\.Experiment statistical significance
32. Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
33. Answer:\[Yes\]
34. Justification: Section 2 and Appendix A\.5 define paired exact McNemar tests, while Sections 3\.1–3\.3 report discordant cells and exact p\-values with an explicit benchmark\-diagnostic interpretation\.
35. Guidelines: - •The answer\[N/A\]means that the paper does not include experiments\. - •The authors should answer\[Yes\]if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper\. - •The factors of variability that the error bars are capturing should be clearly stated \(for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions\)\. - •The method for calculating the error bars should be explained \(closed form formula, call to a library function, bootstrap, etc\.\) - •The assumptions made should be given \(e\.g\., Normally distributed errors\)\. - •It should be clear whether the error bar is the standard deviation or the standard error of the mean\. - •It is OK to report 1\-sigma error bars, but one should state it\. The authors should preferably report a 2\-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified\. - •For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range \(e\.g\., negative error rates\)\. - •If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text\.
36. 8\.Experiments compute resources
37. Question: For each experiment, does the paper provide sufficient information on the computer resources \(type of compute workers, memory, time of execution\) needed to reproduce the experiments?
38. Answer:\[No\]
39. Justification: The models were served by a university Ollama GPU endpoint whose exact accelerator and quantization metadata were not exposed\. Appendix B reports per\-stage call counts, while the archived raw JSONL records include client\-side timings\.
40. Guidelines: - •The answer\[N/A\]means that the paper does not include experiments\. - •The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage\. - •The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute\. - •The paper should disclose whether the full research project required more compute than the experiments reported in the paper \(e\.g\., preliminary or failed experiments that didn’t make it into the paper\)\.
41. 9\.Code of ethics
43. Answer:\[Yes\]
44. Justification: The work uses a public synthetic benchmark, involves no human subjects or private data, and discusses responsible scope and reporting\.
45. Guidelines: - •The answer\[N/A\]means that the authors have not reviewed the NeurIPS Code of Ethics\. - •If the authors answer\[No\], they should explain the special circumstances that require a deviation from the Code of Ethics\. - •The authors should make sure to preserve anonymity \(e\.g\., if there is a special consideration due to laws or regulations in their jurisdiction\)\.
46. 10\.Broader impacts
47. Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
48. Answer:\[Yes\]
49. Justification: The Discussion and conclusion notes that stale sensitive or operational memories can cause outdated recommendations or disclosures, and that version\-aware retirement can mitigate this risk while preserving provenance\.
50. Guidelines: - •The answer\[N/A\]means that there is no societal impact of the work performed\. - •If the authors answer\[N/A\]or\[No\], they should explain why their work has no societal impact or why the paper does not address societal impact\. - •Examples of negative societal impacts include potential malicious or unintended uses \(e\.g\., disinformation, generating fake profiles, surveillance\), fairness considerations \(e\.g\., deployment of technologies that could make decisions that unfairly impact specific groups\), privacy considerations, and security considerations\. - •The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments\. However, if there is a direct path to any negative applications, the authors should point it out\. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation\. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster\. - •The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from \(intentional or unintentional\) misuse of the technology\. - •If there are negative societal impacts, the authors could also discuss possible mitigation strategies \(e\.g\., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML\)\.
51. 11\.Safeguards
52. Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse \(e\.g\., pre\-trained language models, image generators, or scraped datasets\)?
53. Answer:\[N/A\]
54. Justification: The work releases no high\-risk model or newly scraped dataset requiring special safeguards\.
55. Guidelines: - •The answer\[N/A\]means that the paper poses no such risks\. - •Released models that have a high risk for misuse or dual\-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters\. - •Datasets that have been scraped from the Internet could pose safety risks\. The authors should describe how they avoided releasing unsafe images\. - •We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort\.
56. 12\.Licenses for existing assets
57. Question: Are the creators or original owners of assets \(e\.g\., code, data, models\), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
58. Answer:\[Yes\]
59. Justification: Appendix A states that the MemoryAgentBench code repository is MIT licensed and the evaluated Qwen checkpoints are Apache\-2\.0 licensed; the corresponding benchmark and model papers are cited in the main paper\.
60. Guidelines: - •The answer\[N/A\]means that the paper does not use existing assets\. - •The authors should cite the original paper that produced the code package or dataset\. - •The authors should state which version of the asset is used and, if possible, include a URL\. - •The name of the license \(e\.g\., CC\-BY 4\.0\) should be included for each asset\. - •For scraped data from a particular source \(e\.g\., website\), the copyright and terms of service of that source should be provided\. - •If assets are released, the license, copyright information, and terms of use in the package should be provided\. For popular datasets,[paperswithcode\.com/datasets](https://paperswithcode.com/datasets)has curated licenses for some datasets\. Their licensing guide can help determine the license of a dataset\. - •For existing datasets that are re\-packaged, both the original license and the license of the derived asset \(if it has changed\) should be provided\. - •If this information is not available online, the authors are encouraged to reach out to the asset’s creators\.
61. 13\.New assets
62. Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
63. Answer:\[Yes\]
64. Justification: The new scripts, result records, prompts, and audit protocol are documented in the methods and appendices and are released with a README at[https://github\.com/oleeviyababu/Memory\_Conflict](https://github.com/oleeviyababu/Memory_Conflict)\.
65. Guidelines: - •The answer\[N/A\]means that the paper does not release new assets\. - •Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates\. This includes details about training, license, limitations, etc\. - •The paper should discuss whether and how consent was obtained from people whose asset is used\. - •At submission time, remember to anonymize your assets \(if applicable\)\. You can either create an anonymized URL or include an anonymized zip file\.
66. 14\.Crowdsourcing and research with human subjects
67. Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation \(if any\)?
68. Answer:\[N/A\]
69. Justification: The work contains no crowdsourcing or human\-subject data collection\.
70. Guidelines: - •The answer\[N/A\]means that the paper does not involve crowdsourcing nor research with human subjects\. - •Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper\. - •According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector\.
71. 15\.Institutional review board \(IRB\) approvals or equivalent for research with human subjects
72. Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board \(IRB\) approvals \(or an equivalent approval/review based on the requirements of your country or institution\) were obtained?
73. Answer:\[N/A\]
74. Justification: The work contains no human\-subject research and therefore requires no IRB review\.
75. Guidelines: - •The answer\[N/A\]means that the paper does not involve crowdsourcing nor research with human subjects\. - •Depending on the country in which research is conducted, IRB approval \(or equivalent\) may be required for any human subjects research\. If you obtained IRB approval, you should clearly state this in the paper\. - •We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution\. - •For initial submissions, do not include any information that would break anonymity \(if applicable\), such as the institution conducting the review\.
76. 16\.Declaration of LLM usage
77. Question: Does the paper describe the usage of LLMs if it is an important, original, or non\-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does*not*impact the core methodology, scientific rigor, or originality of the research, declaration is not required\.
78. Answer:\[Yes\]
79. Justification: The experimental\-design and supplementary reproducibility sections identify the evaluated LLMs and fully specify how they are used as the systems under study\.
80. Guidelines: - •The answer\[N/A\]means that the core method development in this research does not involve LLMs as any important, original, or non\-standard components\. - •Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described\.