Paper deep dive

Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation

Yufeng Li, Sheng Yao, Peng Zeng, Dezhang Kong, Shi Lin, Changting Lin, Xun Wang, Meng Han

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 103

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Long-term memory supports agents across sessions. Repeated retrieval and write-back can let one historical claim occupy multiple retrieval positions and influence current answers without independent evidence. Existing work examines retrieval, updating, and question answering, leaving same-source record accumulation and post-correction persistence underexplored. Motivated by research in social cognition and memory psychology on repeated exposure, retrieval, and revision, we define \emph{Memory Echo} as the excess retrieval exposure gained by historical information from the same source without additional independent evidence, and introduce MEchoBench to evaluate its emergence, correction resistance, persistence, and effect on answers. To address this phenomenon, we introduce the Family--Provenance--Authority framework (FPA). Specifically, FPA selects one representative per source family, conditions the eligibility of old records for the answer context on verified corrections, and constrains final answers with authoritative evidence. In paired replay on 190 fixed questions across four memory frameworks, FPA raises accuracy from 0.825 to 0.920 with DeepSeek V4 Flash and from 0.782 to 0.916 with Kimi K2.6; mean Decision Contamination Effect falls from 0.106 to 0.030 and from 0.109 to 0.038, respectively. These results show that agent memory can preserve traceable history without allowing repeated historical information to govern current decisions. The code is available at this https URL.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

99,894 characters extracted from source content.

Expand or collapse full text

Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation Yufeng Li 1 * , Sheng Yao 2 * , Peng Zeng 1,† , Dezhang Kong 3 Shi Lin 4 , Changting Lin 3 , Xun Wang 4 , Meng Han 3 1 East China Normal University, Shanghai, China 2 University of Southern California, Los Angeles, USA 3 Zhejiang University, Hangzhou, China 4 Zhejiang Gongshang University, Hangzhou, China Abstract Long-term memory supports agents across ses- sions. Repeated retrieval and write-back can let one historical claim occupy multiple retrieval positions and influence current answers with- out independent evidence. Existing work exam- ines retrieval, updating, and question answer- ing, leaving same-source record accumulation and post-correction persistence underexplored. Motivated by research in social cognition and memory psychology on repeated exposure, re- trieval, and revision, we define Memory Echo as the excess retrieval exposure gained by histori- cal information from the same source without additional independent evidence, and introduce MEchoBench to evaluate its emergence, cor- rection resistance, persistence, and effect on answers. To address this phenomenon, we intro- duce the Family–Provenance–Authority frame- work (FPA). Specifically, FPA selects one rep- resentative per source family, conditions the eligibility of old records for the answer context on verified corrections, and constrains final an- swers with authoritative evidence. In paired re- play on 190 fixed questions across four memory frameworks, FPA raises accuracy from 0.825 to 0.920 with DeepSeek V4 Flash and from 0.782 to 0.916 with Kimi K2.6; mean Decision Contamination Effect falls from 0.106 to 0.030 and from 0.109 to 0.038, respectively. These results show that agent memory can preserve traceable history without allowing repeated his- torical information to govern current decisions. The code is available athttps://github.com /OrcasSam/memory-echo-benchmark. 1 Introduction Long-term memory lets agents carry information across sessions, but current evidence should still de- * Equal contribution. Yufeng Li:yufeng03@stu.ecnu. edu.cn; Sheng Yao:yaosheng@usc.edu. † Corresponding author: Peng Zeng. Email: pzeng@sei.ecnu.edu.cn. User Find keto-friendly, easy-to-prepare, low-carb frozen vegetables. Agent Frozen vegetables Keto bars Continue processing (retrieval / write-back) 1 Recall 1 Keto bars 2 Broccoli mix 3 Frozen vegetables 4 Cauliflower rice 5 Low-carb snacks 2 Update 1 Frozen vegetables 2 Keto bars 3 Cauliflower rice 4 Salad kit 5 Low-carb snacks 3 Re-rank 1 Keto bars 2 Cheese bites 3 Low-carb wraps 4 Almond butter 5 Protein crisps Target: frozen vegetables Final choice: Keto bars High Low FewMany Retrieval / write-back cycles W r o n g - a n s w e r i n f l u e n c e Seed Low influence Reinforce Increasing Correction Influence still high Error persists Decreases slightly SeedReinforceCorrectRe-exposePersist Figure 1: Memory Echo in a shopping query. Repeated write-back gives one historical product several retrieval positions, so its descendants compete with the current correction and can redirect the final choice. termine present answers. A once-suitable product recommendation can become inappropriate, a med- ical answer can be corrected, and a travel itinerary can become obsolete. The issue is not only whether an old record remains stored: repeated retrieval, summarization, and write-back can create multiple retrievable traces from one source. Historical infor- mation can then occupy several positions in later queries and compete with a correction without any new independent evidence. Research on agent long-term memory has ad- vanced along several fronts. A-MEM and Mem0 develop mechanisms for memory organization, re- trieval, and updates (Xu et al., 2025; Chhikara et al., 2025); LoCoMo, LongMemEval, and BEAM evaluate memory question answering over long horizons (Maharana et al., 2024; Wu et al., 2025; Tavakoli et al., 2026). Multi-agent communication and its security risks have been surveyed (Kong et al., 2025); MANBENCH studies collective false memories (Xu et al., 2026a), while corrective re- trieval addresses evidence quality (Yan et al., 2024). These lines of work leave open a lifecycle evalu- ation of how same-source memory accumulates through repeated retrieval and write-back within a single agent, remains exposed after correction, and 1 arXiv:2610.17763v1 [cs.AI] 13 Oct 2026 ultimately affects current answers. Research on socially transmitted memory (Xu et al., 2026a) and on repeated retrieval, recon- solidation, and retention (Roediger and Karpicke, 2006; Nader et al., 2000; Dudai, 2012; Ebbinghaus, 1913) motivates our examination of this process. We call the extra retrieval exposure that histori- cal information from one source gains through re- peated processing without additional independent evidence Memory Echo. In a shopping pilot, re- peated read-only retrieval leaves the recorded task scores unchanged, whereas writing interactions back increases historical occupancy and reduces an- swer quality. Final-answer accuracy alone cannot distinguish whether an error arises from trace pro- liferation, a correction failing to displace old traces, or a generator failing to use a retrieved correction. We organize the study around four questions. RQ1 (formation): Across four memory architec- tures, how do seed-only storage, repeated read-only retrieval, and retrieval with write-back differ in same-source derivative count, historical Echo Den- sity (ED), and old-answer adoption? RQ2 (post- correction influence): After a common verified correction, what remains in native retrieval versus served context, how does exposure persist through 30 read-only queries, and does it still lower answer accuracy? RQ3 (FPA mechanism): With matched candidates and a fixed generator, how much do Family, Provenance, and Authority each contribute beyond Baseline, Budget, CRAG-M, and PairLR, including the change from unconstrained to con- strained answers? RQ4 (transfer and robustness): Do paired gains persist across four architectures, two generators, and shopping, medical, and travel tasks under provenance errors and incomplete cor- rections, while retaining valid history? MEchoBench pairs source questions across memory histories and follows initialization, repeti- tion, correction, delayed retrieval, and task answer- ing. We track derivative counts and ED for RQ1; native and served retrieval, storage- and serving- side Reconsolidation Resistance Index (RRI), a 30-query ED curve, Windowed Echo Persistence (WEP), and final accuracy for RQ2; Decision Con- tamination Effect (DCE) and before/after answer changes for RQ3; and paired performance and valid-history retention for RQ4. The formation and correction mechanisms in RQ1 and RQ2 motivate the Family–Provenance– Authority framework (FPA), and RQ3 asks which of its components account for the measured gains. FPA selects one retrieval representative per source family to limit the occupancy advantage of repeated traces, uses verified corrections to determine when old records may enter the current answer context, and applies authoritative evidence and answer con- straints to reduce historical pull on the final output. Old records remain in storage, allowing storage and serving effects to be observed separately; RQ4 asks whether this balance survives changes in ar- chitecture, generator, task, and provenance quality. We evaluate the memory lifecycle across four general data sources and compare FPA with three mechanisms on matched candidate pools. On 190 fixed source questions across four architectures, FPA raises DeepSeek accuracy from 0.825 to 0.920 and Kimi accuracy from 0.782 to 0.916; mean DCE falls from 0.106 to 0.030 and from 0.109 to 0.038, respectively. Shopping, medical question answer- ing, and travel planning show how retrieval changes affect current answers. Our contributions are: •We introduce Memory Echo as a source-level phenomenon and characterize how repeated processing gives historical information a re- trieval advantage that can affect decisions. •We develop MEchoBench and four measure- ments to track trace accumulation, correction resistance, delayed persistence, and answer shift within one memory lifecycle. •We introduce FPA, combining family repre- sentative selection, post-correction eligibility, and authority-guided answering to reduce de- cision contamination while preserving trace- able history, and validate its effects across general data sources and practical tasks. 2 Related Work 2.1 Long-Term Memory Evaluation External memory gives an agent access to infor- mation beyond its prompt. A-MEM organizes linked notes, while Mem0 supports storing, search- ing, and updating memories (Xu et al., 2025; Chhikara et al., 2025). LoCoMo, LongMemEval, and BEAM test long-history question answering (Maharana et al., 2024; Wu et al., 2025; Tavakoli et al., 2026); HotpotQA tests evidence combination (Yang et al., 2018). MemoryAgentBench addition- ally evaluates incremental learning and selective forgetting (Hu et al., 2026). Li et al. (2026) study whether retrieved memories jointly contain suffi- cient evidence to answer a question. Echo asks 2 whether repeated processing makes one source ap- pear to be several independent pieces of evidence. Among industry-backed agent-memory systems, LangChain’s LangMem is a prominent open-source project that provides tools for memory creation, search, and background consolidation (LangChain Team, 2025). Work on retrieval practice and memory reten- tion motivates our sequence of access, revision, and delayed observation (Roediger and Karpicke, 2006; Nader et al., 2000; Dudai, 2012; Ebbinghaus, 1913). Here these are software events: searches, writes, and later retrievals. Corrective RAG as- sesses evidence quality (Yan et al., 2024), while learning-to-rank changes candidate order (Burges et al., 2005). Our CRAG-M comparator grades stored memories, and PairLR uses a linear pairwise logistic ranker. Neither identifies candidates de- scended from the same claim; Echo and FPA track that shared ancestry. 3 Preliminaries MotivationInshopping-0028, the user requests low-carb frozen vegetables. Write-back multiplies an older keto-bar recommendation; its descendants crowd the correction and redirect the choice (Fig- ure 1). Read-only search leaves storage unchanged, isolating the effect of write-back. Agent Memory and Evidence Provenance At turnt, retrieverRsearches memoryM t for query q t ; serving ruleGchooses the answer context, and updaterUchanges storage. Each record carries content, time, source and family IDs, a parent pointer, validity, and exposure counts. A retrieval- derived paraphrase inherits its parent’s source: sev- eral records can be one piece of evidence. Operational Definition of Memory EchoMem- ory Echo is the extra retrieval exposure a historical claim and its descendants gain without independent evidence. The original claim need not have been false: the failure occurs when it keeps occupying evidence positions after the task changes or a ver- ified correction arrives. We separate record from source counts and retrieval from answer bias. 4 Agent Memory Echo One source becomes several records; correction fails to displace them; exposure persists into an answer. RQ1–RQ3 trace this lifecycle; RQ4 tests transfer across architectures, generators, and tasks. 4.1 The MEchoBench Lifecycle Questions and historical seeds. We take 50 questions each from LoCoMo, LongMemEval, Hot- potQA, and BEAM (Maharana et al., 2024; Wu et al., 2025; Yang et al., 2018; Tavakoli et al., 2026), reserving ten LoCoMo questions for development and reporting 190. We retain wording, gold answer, supporting evidence, ability type, and source ID. A different real answer from the same dataset, prefer- ably matching ability type, seeds history; donor IDs and derivative parent links preserve ancestry. The target gold answer supplies the verified correction. Paired histories and five stages.For each ques- tion, three histories share the seed, neutral com- petitors, query sequence, correction, and retrieval budget. Once stores the seed once; Read retrieves it six times without writing; Write follows each retrieval with a same-family derivative write. En- coding records the initial state, retrieval practice changes exposure, and updating inserts the same correction in all histories. Thirty subsequent read- only probes measure persistence without further writes; final retrieval supplies the answer context. Figure 2 maps these five stages to observations, with probe locations in Appendix Table 5. Lifecycle readouts. At encoding, we measure how much of the initial retrieval is occupied by the seed. After retrieval practice, we count its descen- dants and their share of the pre-correction retrieval. Once the correction is added, we check whether it is retrieved and how much of the old family re- mains in storage and in the returned list. We then run 30 read-only probes, recording retrieval scores without changing memory. Finally, we record the answer produced from the candidates served to the generator. These measurements distinguish an old claim that remains stored from one that is retrieved, and one that the answer actually adopts. What remains matched.We keep the final query and gold answer fixed across comparisons. The system first forms its native memory and retrieves candidates; we save both before replaying each serving method on the same candidate pool. This separates the records the system creates from the evidence the generator actually sees. Questions from the general datasets test source-level Echo, while the public-data shopping, medical-QA, and travel groups test whether it carries over to current tasks (Yao et al., 2022; Jin et al., 2021; Xie et al., 2024). The experiments cover four memory ar- 3 EncodingRetrieval Practice Updating / Reconsolidation Delayed RetrievalFinal Decision ED ED t = k ∑ i=1 w i s i,t 퐀 (m i,t ∈ ℱ e ) k ∑ i=1 w i w i = 1 log 2 (i+1) s i,t =max ( 0,cos(m i,t ,e) ) RRI RRI ℓ = 1 2 min1, E + ℓ E − ℓ + 1 2 E + ℓ E + ℓ +C + ℓ RRI * = RRI storage +RRI serving 2 WEP WEP 30 = 1 29 29 ∑ t=1 ED t +ED t+1 2 ED Session 1–30 DCE DCE(a)= d(a,g) d(a,g)+d(a,e) d(x,y)=1−cos(x,y) Figure 2: MEchoBench’s five stages.E-minus/E- plus mark old exposure before/after correction; C- plus marks corrected exposure. Here a/g/e denote an- swer/gold/Echo texts, and d is cosine distance. ED, RRI, WEP, and DCE use Equations 1–4; the pictured RRI and DCE formulas are schematic. chitectures and two answer generators (Section 6). Because the memory trajectory stays fixed during replay, a change in the served context is not mis- taken for a change in native memory. 4.2 Metric Definitions Echo Density (ED). ED measures how much of the ranked answer context is occupied by one his- torical source family: ED = P N r=1 w r b r s r P N r=1 w r .(1) Symbols.Nis the number of retrieved records; ris a record’s rank from1(highest) toN.b r is 1if the record at rankrdescends from the his- torical seed and0otherwise.x r andeare se- mantic representations of that record and the seed; s r = max(0, cos(x r ,e))is their nonnegative co- sine similarity.w r = 1/ log 2 (r + 1)is the rank weight, withlog 2 denoting the base-two logarithm. The sums add contributions over all ranks. Thus the numerator counts similarity-weighted old-family positions, and the denominator normalizes by all rank weights. High ED means prominent exposure, not necessarily a wrong answer. Reconsolidation Resistance Index (RRI). RRI asks whether a verified correction displaces the old family. It combines the family’s survival with its competition against the corrected evidence: RRI = 1 2 min(1, ED post /ED pre ) + 1 2 ED post ED post + C post . (2) Symbols.ED pre andED post are old-family oc- cupancies before and after correction;C post is corrected-family occupancy afterward. Retrieval RRI uses rank-weighted occupancies of returned records; storage RRI uses unranked occupancies of stored records. Each family’s semantic similar- ity is measured against its own seed or correction. min(1,·)caps survival at one; the second fraction is the old share of competing evidence. The two 1/2factors give both parts equal weight, yielding RRI in[0, 1]. RRI is zero whenED pre = 0; a zero denominator in the second fraction also makes that fraction zero. Windowed Echo Persistence (WEP). WEP summarizes old-family exposure across the con- secutive read-only queries after correction: WEP 30 = 1 29 29 X t=1 ED t + ED t+1 2 .(3) Symbols.ED t is old-family ED at read-only probe t, andED t+1 is its value at the next probe. There are 30 probes; the sum runs overt = 1,..., 29 to cover their 29 adjacent intervals. The index counts queries, not elapsed time. Each adjacent pair is averaged by dividing by2, and division by29 averages across the full window.WEP 30 therefore measures persistence even when ED fluctuates. No probe writes a record. Decision Contamination Effect (DCE). DCE tests whether the generated answer aligns more with the historical claim than with the verified an- swer: DCE = max(0, (E− G)/4).(4) Symbols.ais the generated answer.E,G∈ [0, 4] are its judge-assigned alignment scores with the historical Echo and verified gold answers, respec- tively; zero means no alignment and four means full alignment. DividingE − Gby4normalizes their difference, andmax(0,·)clips negative val- ues to zero. ThusDCE ∈ [0, 1]measures the an- swer’s extra alignment with the old claim. Reading the metrics together. ED locates old- family occupancy, RRI its resistance to correction, WEP its delayed exposure, and DCE its influence on the answer. Reporting each at its stage sepa- rates a crowded retrieval list from a contaminated decision. Correctness is judged against the verified answer, without showing the Echo, on a 0–4 scale. Scores of at least 3 count as accurate; a separate binary judgment records old-answer adoption. 4 5 Provenance-Guided Mitigation 5.1 The Three-Stage FPA Pipeline FPA controls the answer context between retrieval and generation without changing stored history. Records carry source and family IDs, parent links, source type, and validity; derivatives inherit their parent’s family. Figure 3 shows family selection, provenance gating, and authority-guided answering in a shopping query. Stage 1: Family-level selection.Source and par- ent links group the old whole-milk records M1– M3 in Figure 3 into family F1; M4–M6 have other origins. FPA scores retrieval relevance, con- fidence, source reliability, exposure, and validity (Appendix E), then admits at most one represen- tative per family. M1 represents F1 in the illustra- tion, while M2 and M3 receive no extra positions. Repeated writes therefore add records, not inde- pendent votes. The family ID follows ancestry, not the wording of the answer: two independent sources may agree, while two differently worded summaries of one retrieval remain the same fam- ily. The representative is selected anew from the current candidate pool; source grouping does not delete the unselected records. Stage 2: Provenance-aware gating. The gate checks source, parent, verification, and validity against the correction. In the figure, M4 supersedes the whole-milk family: M1–M3 remain stored for audit but cannot support the current answer. M4 enters as verified evidence; nonconflicting M5 and M6 remain eligible context. Retrieval nei- ther grants a record write-back authority nor re- stores an obsolete descendant’s evidential status. Section 6.6 tests this gate when correction quality changes. This is the handoff from family balance to evidence validity. Selecting only M1 from F1 would still leave one obsolete whole-milk recom- mendation in the prompt; the provenance check removes that last route into current evidence after M4 arrives. Serving eligibility changes while the full record lineage remains available. Stage 3: Authority-guided answering.The gen- erator sees eligible evidence rather than raw Top-k and is directed to prefer verified current facts. Here M4 supports “You should buy oat milk,” not the superseded whole-milk claim. A final check com- pares the draft with admitted authority and may select a supported short span for an extractive ques- tion; both raw and constrained answers are retained. The figure’s no-authority fallback to qualified con- text is illustrative. The missing/noisy-correction experiment tests the gate without implementing that recovery branch. Serving comparison.All methods serve at most five records from the same frozen pool. Budget penalizes excess old occupancy; CRAG-M grades candidates (Yan et al., 2024); PairLR learns pair- wise ranking (Burges et al., 2005). FPA addition- ally controls family positions, post-correction eligi- bility, and the final answer’s authority. Appendix E gives the protocols. 6 Experiments 6.1 Experimental Setup The lifecycle runs in A-MEM (Xu et al., 2025), official Mem0 (Chhikara et al., 2025), CoM (Xu et al., 2026b), and LangMem (LangChain Team, 2025). Two reinforced histories for each of 190 questions in four architectures yield 1,520 paired answer states per method and generator. DeepSeek V4 Flash uses temperature 0 and Kimi K2.6 uses 0.6; both produce at most 120 tokens.GPT- 5.6 (gpt-5.6-sol) judges answers in two passes. Within each generator, methods share questions, histories, and candidate pools. ED and WEP de- pend on the served pool, so they match across gen- erators. We compare methods within each model; Appendix B reports architecture cells. 6.2 Metrics Accuracy uses the judge’s correctness score, while old-answer rate records whether the generated an- swer adopts the historical claim. ED, RRI, WEP, and DCE retain the distinct roles defined in Sec- tion 4. Methods are compared on paired question– architecture–history states. For uncertainty inter- vals, we first average states belonging to the same source question and then resample the 190 ques- tions, so repeated states from one question do not become independent observations. 6.3 Main Results The lifecycle experiments locate where Echo enters the pipeline, and the main method comparison asks whether that influence can be interrupted when the source question, correction, and candidate pool remain fixed. Table 2 reports both answer models; the stage-specific contrasts explain its results. 5 Stage 1 — F: Family-level Selection What should I buy? User Question External Memory Store Retrieved Candidate Memories M1 Buy whole milk. M2 Whole milk is on the list. M3 Remember to buy whole milk. M4 Buy oat milk instead of whole milk. M5 I saved a coupon. M6 The store moved last week. Family F1 — Shared origin Derived fromDerived from M1 Buy whole milk. M2 Whole milk is on the list. M3 Remember to buy whole milk. Family F2 M4 Buy oat milk instead of whole milk. Family F3 M5 I saved a coupon. Family F4 M6 The store moved last week. Select one representative per family Relevance Source reliability Current status M1 Buy whole milk. (F1 representative) M2 Whole milk is on the list. (not selected) M3 Remember to buy whole milk. (not selected) M4 Buy oat milk instead of whole milk. (F2 representative) M5 I saved a coupon. (F3 representative) M6 The store moved last week. (F4 representative) Stage 2 — P: Provenance-aware Gating Input Candidates M1 Buy whole milk. (F1) M4 Buy oat milk instead of whole milk. (F2) M5 I saved a coupon. M6 The store moved last week. (F4) Check provenance & status Source · Parent record · Status Independent source? Verified or user-confirmed? Still valid? Gating Results Historical / Retrieval-derived (unverified, outdated) M1 Buy whole milk. Superseded M2 Whole milk is on the list. Superseded M3 Remember to buy whole milk. Superseded Retained for audit M1 M2M3 (Not eligible for answering) Eligible Evidence M4 Buy oat milk instead of whole milk. Verified correction M5 I saved a coupon. Valid context M6 The store moved last week. Stage 3 — A: Authority-guided Answering Eligible Evidence M4 Buy oat milk instead of whole milk. Verified correction M5 I saved a coupon. M6 The store moved last week. Authoritative Context Prefer verified current evidence; otherwise use qualified user-confirmed context. No authoritative evidence Fallback context What should I buy? User Question Agent Answer from current authoritative evidence. Answer Constraint Extract a directly supported short answer when applicable. Final Answer You should buy oat milk. Whole milk (Not allowed) Retrieve Family Grouping Family-balanced Candidates Corrects Figure 3: Three-stage FPA pipeline: one serving position per source family, provenance-based eligibility, and an answer grounded in current authority. Superseded records remain auditable; the fallback box is illustrative, not an evaluated recovery rule. RQ1: Same-Source Retrieval Occupancy The first contrast separates remembering a claim from multiplying it. Repeated read-only search leaves pre-correction ED essentially unchanged from one-time encoding.When the same searches write derivatives back, old-family occupancy rises sharply in every architecture, and old answers reap- pear after the common correction (Appendix Ta- ble 6). The extra records all inherit one source: write-back amplifies access to a claim without adding independent support. Its cross-architecture consistency points to the retrieval–write feedback loop, not a particular ranker. RQ2: Correction Resistance and Persistence All histories receive the same verified correction, yet write-back lets old descendants crowd it out of retrieval. Storage and retrieval RRI rise as descen- dants survive the update and compete for limited evidence slots. Appendix A shows this competition is strongest in CoM. Stopping writes does not erase that advantage. The read-only ED values remain elevated and nearly flat across the delayed probes (Table 1); WEP therefore captures an inherited retrieval bias rather than ongoing reinforcement. The complete trajectories are in Appendix Tables 7–8. The ver- ified answer is stored while descendants of the historical answer still fill the serving path (Ap- pendix I). The ability-level radar shows that this exposure and its effect on answers vary by ques- 0.5 0.7 ED 0.4 0.6 RRI Base Budget CRAG-M PairLR FPA 0.1 0.4 WEP Base Budget CRAG-M PairLR FPA 0.04 0.10 DCE Figure 4: DeepSeek method profiles for ED, RRI, WEP, and DCE; lower is better. ED RRI WEP DCE 0.5 1.0 ED RRI WEP DCE 0.5 1.0 Adversarial Multi-hop Open-dom. Single-hop Temporal Update Multi-sess. Asst. recall User recall Temporal (a) LoCoMo(b) LongMemEval Figure 5: Echo by ability in LoCoMo and Long- MemEval. Exact values appear in Appendix Table 15. tion type, so ED alone cannot stand in for decision contamination (Figure 5; Appendix Table 15). RQ3: Answers to Current Tasks The lifecycle establishes how an old family reaches the generator; Table 2 asks whether serving can interrupt that path. FPA improves accuracy and reduces old- answer adoption with both DeepSeek and Kimi. Their ED and WEP changes coincide because the intervention alters the shared evidence pool before 6 MemoryPrePostt=1t=10t=20t=30OldRet. A-MEM (Xu et al., 2025)0.9320.7040.6600.6640.6650.6540.1951.000 Mem0 (Chhikara et al., 2025)0.9320.7040.6600.6640.6650.6540.2111.000 CoM (Xu et al., 2026b)0.9320.7220.6810.6870.6840.6710.2321.000 LangMem (LangChain Team, 2025)0.9320.7040.6600.6640.6650.6540.2111.000 Table 1: DeepSeek persistence after write-back (190 questions per row). All 30 delay queries were read-only. Pre, Post, and the four query columns report ED; Old is the final old-answer rate and Ret. is the final old-retrieval proxy. either model generates an answer. DCE also falls in both blocks, connecting lower exposure to a weaker pull on the final response. Budget, CRAG-M, and PairLR often improve accuracy without removing persistent old-family exposure. CRAG-M is especially revealing: a use- ful candidate can help an answer even while same- source descendants still occupy ranked positions. FPA instead changes both how many positions a family receives and whether its obsolete mem- bers remain eligible after correction. The delayed- exposure contrast makes this mechanism clearer than a single search snapshot. 6.4 Framework and Model Variation The architecture breakdown reveals two regimes (Appendix B). A-MEM and Mem0 show large accuracy gains alongside near-elimination of de- layed exposure. CoM and LangMem improve more modestly in accuracy, yet their WEP also falls: a crowded final retrieval does not imply sustained ex- posure across the delay window. Kimi follows the same broad split (Appendix Table 10). Question- level paired bootstrap intervals support positive accuracy changes and lower DCE for both genera- tors. We resample source questions rather than their correlated architecture–history states, so the uncer- tainty reflects variation across questions rather than repeated views of the same one. Figure 6 makes the architecture split visible relative to each Base- line. A-MEM and Mem0 nearly eliminate persis- tent exposure and decision contamination. CoM and LangMem show a different response: the final ED can remain close to Baseline while WEP and DCE fall. The intervention therefore changes re- peated exposure and answer influence even where a single retrieval snapshot changes little. Appendix Figure 8 compares shared context changes with the two generators’ answer responses. 6.5 Paired Mechanism Comparisons The matched comparisons show that FPA couples answer improvement with a larger drop in persis- Old EDWEPDCE 0.00 0.25 0.50 0.75 1.00 Fraction of baseline 0.00 0.57 0.01 0.00 A-MEM Old EDWEPDCE 0.00 0.60 0.01 0.00 Mem0 Old EDWEPDCE 0.00 0.25 0.50 0.75 1.00 Fraction of baseline 0.55 1.01 0.32 0.57 CoM Old EDWEPDCE 0.48 1.01 0.20 0.50 LangMem Figure 6: DeepSeek V4 Flash risk under FPA relative to each architecture’s Baseline (Table 10). The dashed line marks Baseline= 1; shorter bars indicate lower risk on the displayed metrics. MethodAcc.↑Old↓ED↓WEP↓DCE↓ DeepSeek V4 Flash Baseline0.8250.1060.6620.4270.106 Budget0.8610.0590.5740.2640.056 CRAG-M0.8550.0590.7070.2450.055 PairLR0.8610.0580.6920.3090.054 FPA0.9200.0280.5280.0570.030 Kimi K2.6 Baseline0.7820.1080.6620.4270.109 Budget0.8470.0750.5740.2640.073 CRAG-M0.8420.0640.7070.2450.064 PairLR0.8450.0680.6920.3090.065 FPA0.9160.0380.5280.0570.038 Table 2: Main comparison. Four-architecture means over 1,520 paired states per generator. Accuracy is higher-is-better; other metrics are lower-is-better. ED and WEP depend on the served pool. CRAG-M adapts Yan et al. (2024); PairLR follows Burges et al. (2005). Frameworks: A-MEM (Xu et al., 2025), Mem0 (Chhikara et al., 2025), CoM (Xu et al., 2026b), LangMem (LangChain Team, 2025). tent exposure than the three alternative serving rules. Kimi’s paired accuracy gains are positive across all four frameworks (Appendix Table 9), while its DCE ordering is unchanged. The paired in- tervals in Appendix Figure 10 separate this family- level exposure effect from a generic improvement 7 00.51 Shopping Medical QA Travel (a) Post ED 00.51 (b) RRI 00.51 Shopping Medical QA Travel (c) WEP 00.51 (d) DCE BaselineFPA Figure 7: Legacy three-scenario pilot. Rows show shop- ping, medical QA, and travel; circles are Baseline and diamonds are FPA. Exact values appear in Appendix Table 16. The pilot used an earlier extractive protocol, not the current task-transfer cohort. in answer selection. CRAG-M and PairLR can elevate useful candidates without excluding same- source descendants. FPA changes eligibility be- fore ranking, so its WEP advantage persists across the read-only delay. The paired intervals resample source questions, preserving the shared comparison across architectures and histories. 6.6 Variation across Data Sources Task transfer depends on whether history actually conflicts with current evidence. In shopping and medical QA, FPA reduces old-answer adoption while improving accuracy. Travel begins without old-answer adoption; restricting useful planning context then reduces accuracy (Table 12). The result identifies a boundary of the intervention, not a universal benefit from filtering history. An earlier extractive pilot measured Post ED, RRI, WEP, and DCE on shopping, medical QA, and travel (Figure 7). It used 50 tasks per sce- nario across A-MEM and Mem0. Its values are in Appendix Table 16; this pilot is distinct from the current transfer study (Table 12). Source-quality tests sharpen that boundary. Complete provenance and noisy family IDs retain a modest advantage, whereas a missing or unreliable correction leaves the gate without a trustworthy re- placement and harms accuracy. When the old fact is nonconflicting, the change is small (Appendix Table 13). The contrast separates evidence validity from recency: filtering helps only when reliable current evidence replaces the old source. VariantAcc.OldDCE Baseline0.8250.1060.106 w Family0.6510.0950.097 w Provenance0.8200.0100.009 w Joint gate0.6370.0120.011 w Authority0.9120.0310.032 Full FPA0.9200.0280.030 Table 3: DeepSeek component ablation on 1,520 paired states. The full component-presence matrix is in Ap- pendix Table 11. 6.7 Ablation Results: Sequential Additions The ablation separates suppressing obsolete evi- dence from supplying a usable answer. Provenance control alone sharply reduces old-answer adoption but does not improve correctness; family control by itself can also remove useful context. Combining the two gates further reduces the historical claim’s reach but leaves the generator short of a replace- ment answer. Accuracy recovers when verified au- thority is added, and the answer check contributes a smaller final gain (Table 3). Together, these stages limit unsupported repetition and supply the current fact. 6.8 Representative Success and Failure Untreated DeepSeek repeats the historical answer, whereas FPA returns the verified one. The trace links that reversal to which source family reaches the generator, not to a changed question (Sec- tion 6.3). Travel and unreliable-correction cases show the converse: when the old answer is not be- ing adopted or replacement evidence is unavailable, withholding context can remove information the generator still needs. The serving decision must therefore depend on a real conflict and a usable authority; Table 4 illustrates the correction-quality boundary. Source conditionBaselineFPA Complete0.6040.629 Correction missing/noisy0.3780.212 Table 4: DeepSeek accuracy by correction quality; 760 answer states per method and condition. Full results appear in Appendix Table 13. 7 Conclusion We identify Memory Echo as a failure mode in ex- ternal agent memory, where repeated write-back turns a single historical source into seemingly inde- pendent evidence. Across four architectures, this 8 effect persists after correction and 30 unreinforced queries, but does not arise from read-only access. MEchoBench traces its propagation from source ancestry through retrieval to answers. Our findings distinguish memory-side exposure from generator-dependent outcomes and reveal lim- itations of final retrieval snapshots. At the serving boundary, FPA reduces Echo exposure and im- proves accuracy across two generators without dis- carding history. These results highlight a central principle for reliable agent memory: preserving historical information should not grant repeated records independent evidential authority, particu- larly after verified correction. Limitations The benchmark places a real answer from another question into a controlled historical position; it does not estimate how frequently such seeds arise in natural interaction. Serving-method compar- isons replay recorded candidate pools rather than evolving a separate memory database for each com- parator. FPA uses source-family and validity anno- tations supplied by the benchmark, and its final an- swer check can select a concise verified span, mak- ing part of the measured gain an answer-processing effect. GPT-5.6 judges correctness and contamina- tion without an independent human audit of bor- derline cases. The travel and missing-correction conditions show that the measured benefit depends on a relevant Echo conflict and usable replacement evidence. Ethical Considerations Medical questions are used as evaluation cases, not clinical advice. Historical seeds intentionally con- flict with verified answers and belong in controlled benchmark traces. Source questions, derived tra- jectories, and logs should be redistributed in ac- cordance with each dataset’s license and privacy conditions. References Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. 2005. Learning to rank using gradient descent. In Proceedings of the 22nd International Conference on Machine Learning, pages 89–96. Association for Computing Machinery. Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. Mem0: Building production-ready AI agents with scalable long-term memory. In Proceedings of the 28th European Con- ference on Artificial Intelligence, volume 413 of Fron- tiers in Artificial Intelligence and Applications, pages 2993–3000. IOS Press. Yadin Dudai. 2012. The restless engram: Consolida- tions never end. Annual Review of Neuroscience, 35:227–247. Hermann Ebbinghaus. 1913. Memory: A Contribu- tion to Experimental Psychology. Teachers Col- lege, Columbia University, New York. Translated by Henry A. Ruger and Clara E. Bussenius; original German edition published in 1885. Yuanzhe Hu, Yu Wang, and Julian McAuley. 2026. Eval- uating memory in LLM agents via incremental multi- turn interactions. In The Fourteenth International Conference on Learning Representations. Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2021. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Ap- plied Sciences, 11(14):6421. Dezhang Kong, Shi Lin, Zhenhua Xu, Zhebo Wang, Minghao Li, Yufeng Li, Yilun Zhang, Hujin Peng, Xiang Chen, Zeyang Sha, Yuyuan Li, Changt- ing Lin, Xun Wang, Xuan Liu, Ningyu Zhang, Chaochao Chen, Chunming Wu, Muhammad Khur- ram Khan, and Meng Han. 2025. A survey of LLM- driven AI agent communication: Protocols, security risks, and defense countermeasures. arXiv preprint arXiv:2506.19676. LangChain Team. 2025. LangMem: Agent memory management tools. GitHub repository. Accessed 10 October 2026. Yufeng Li, Shuxin Li, Zhenhua Xu, Junxian Li, Peng Zeng, Sheng Yao, Changting Lin, Gaolei Li, Ran Bi, and Meng Han. 2026. Relevance is not sufficiency: What actually closes the evidence gap in long-term memory QA. arXiv preprint arXiv:2610.09348. Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. 2024.Evaluating very long-term conversational memory of LLM agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13851– 13870, Bangkok, Thailand. Association for Compu- tational Linguistics. Karim Nader, Glenn E. Schafe, and Joseph E. Le Doux. 2000. Fear memories require protein synthesis in the amygdala for reconsolidation after retrieval. Nature, 406(6797):722–726. Henry L. Roediger, I and Jeffrey D. Karpicke. 2006. Test-enhanced learning: Taking memory tests im- proves long-term retention. Psychological Science, 17(3):249–255. 9 Mohammad Tavakoli, Alireza Salemi, Carrie Ye, Mo- hamed Abdalla, Hamed Zamani, and J Ross Mitchell. 2026. Beyond a million tokens: Benchmarking and enhancing long-term memory in LLMs. In The Four- teenth International Conference on Learning Repre- sentations. Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai- Wei Chang, and Dong Yu. 2025. LongMemEval: Benchmarking chat assistants on long-term interac- tive memory. In The Thirteenth International Con- ference on Learning Representations. Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. 2024. TravelPlanner: A benchmark for real-world planning with language agents. In Proceedings of the 41st International Conference on Machine Learning, vol- ume 235 of Proceedings of Machine Learning Re- search, pages 54590–54613. PMLR. Naen Xu, Hengyu An, Shuo Shi, Jinghuai Zhang, Chunyi Zhou, Changjiang Li, Tianyu Du, Zhihui Fu, Jun Wang, and Shouling Ji. 2026a. When agents “misremember” collectively: Exploring the Mandela effect in LLM-based multi-agent systems. In The Fourteenth International Conference on Learning Representations. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2025. A-Mem: Agentic memory for LLM agents. In Advances in Neural Information Processing Systems, volume 38, pages 17577–17604. Xiucheng Xu, Bingbing Xu, Tian Xueyun, Zihe Huang, Rongxin Chen, Li Yunfan, and Huawei Shen. 2026b. Chain-of-memory: Lightweight memory construc- tion with dynamic evolution for LLM agents. In Proceedings of the 64th Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 11618–11631, San Diego, Cali- fornia, United States. Association for Computational Linguistics. Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, and Zhen-Hua Ling. 2024. Corrective retrieval augmented generation. arXiv preprint arXiv:2401.15884. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christo- pher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing, pages 2369–2380, Brussels, Belgium. Association for Com- putational Linguistics. Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. 2022. WebShop: Towards scalable real- world web interaction with grounded language agents. In Advances in Neural Information Processing Sys- tems, volume 35, pages 20744–20757. 10 Appendix Contents A. Lifecycle (Tables A–C) . . . . . . . . . . . . . . . . . . . 11 B. Main methods (Table D; Kimi gains) . . . . . . . 11 C. Ablations and Kimi (E–F) . . . . . . . . . . . . . . . . 12 D. Task and source (G–H) . . . . . . . . . . . . . . . . . . . 12 E. Reproducibility . . . . . . . . . . . . . . . . . . . . . . . . . . 12 F. Lifecycle interpretation . . . . . . . . . . . . . . . . . . . 14 G. Comparator protocols . . . . . . . . . . . . . . . . . . . . 14 H. Supplementary visual comparisons . . . . . . . . 15 I. Five-stage cases (25) . . . . . . . . . . . . . . . . . . . . . . 16 I.1 Encoding I.2 Practice I.3 Update I.4 Delay I.5 Decision The entries above link to the all appendix sec- tions, including the refreshed case catalogue. The Overleaf package contains only sources needed to compile the paper; machine-readable experiment records are maintained separately. A Full lifecycle measurements The full condition breakdown separates three mo- ments in the same Echo trajectory. Table 6 shows how each history changes occupancy before correc- tion and old-answer adoption afterward. Table 7 locates the correction in storage and retrieval. Ta- ble 1 samples the subsequent read-only curve; WEP includes every one of its 30 probes. The Once and Read histories provide a direct control for exposure without additional records: both have the same ED and no derived write. The Write history adds six same-source records before the common correction. Its post-correction old- answer rates range from 0.195 to 0.232, so the retrieval change has an observable answer conse- quence. The storage RRI in Table 7 then shows that correcting the original record does not remove its descendants; retrieval RRI reports how many still reach the ranked output. B Main experiment by architecture The pooled result conceals different starting points across memory systems. Table 10 places DeepSeek and Kimi side by side and retains all architecture– method cells, each averaging 380 answer states. Accuracy and old-answer rate come from gener- ated text; ED and WEP describe the served mem- ory. For Kimi, paired FPA minus Baseline ac- curacy gains are 0.182 [0.145,0.218] on A-MEM, 0.179 [0.142,0.216] on Mem0, 0.097 [0.066,0.132] on CoM, and 0.082 [0.050,0.116] on LangMem StageProbe purposeQA count EncodingIs the historical seed re- trievable? 1 Retrieval practice Six accesses plus one com- mon pre-correction probe 7 Updating Does verified evidence dis- place the seed? 1 Delayed retrievalDoes old exposure remain after writes stop? 30 Final decisionDoes the current answer follow old or new evi- dence? 1 Table 5: Stage-specific probes; delay queries are con- secutive and read-only. Source questions: LoCoMo (Maharana et al., 2024), LongMemEval (Wu et al., 2025), HotpotQA (Yang et al., 2018), BEAM (Tavakoli et al., 2026). MemoryHistoryED Derivatives Old rateRet. A-MEM 1 Once0.39000.000 0.505 Read0.39000.000 0.505 Write0.93260.195 1.000 Mem0 2 Once0.40200.000 0.505 Read0.40200.000 0.505 Write0.93260.211 1.000 CoM 3 Once0.39000.000 0.505 Read0.39000.000 0.505 Write0.93260.232 1.000 LangMem 4 Once0.39000.000 0.505 Read0.39000.000 0.505 Write0.93260.211 1.000 Table 6: DeepSeek formation results (190 questions per row). ED is measured before correction; Derivatives counts write-back records, Old is the post-correction old-answer rate, and Ret. is the final old-retrieval proxy. 1 A-MEM (Xu et al., 2025); 2 Mem0 (Chhikara et al., 2025); 3 CoM (Xu et al., 2026b); 4 LangMem (LangChain Team, 2025). MemoryCorr. RRI s RRI r WEP 30 Acc. A-MEM 1 O/R1.000 0.519 0.4540.190 0.984 A-MEM 1 Write0.797 0.872 0.7630.658 0.684 Mem0 2 O/R1.000 0.519 0.4510.195 0.984 Mem0 2 Write0.797 0.872 0.7630.658 0.663 CoM 3 O/R1.000 0.519 0.4540.190 0.984 CoM 3 Write0.688 0.872 0.7850.677 0.642 LangMem 4 O/R1.000 0.519 0.4540.190 0.984 LangMem 4 Write0.797 0.872 0.7630.658 0.674 Table 7: DeepSeek correction results. O/R combines the identical Once and Read outcomes;sandrdenote storage and retrieval RRI. Acc. is final-answer accuracy. 1 A-MEM (Xu et al., 2025); 2 Mem0 (Chhikara et al., 2025); 3 CoM (Xu et al., 2026b); 4 LangMem (LangChain Team, 2025). (95% source-level bootstrap intervals over 190 questions). The architecture rows distinguish a change in final retrieval rank from a change in accumulated exposure. On CoM and LangMem, FPA’s final ED 11 QueryA-MEM 1 Mem0 2 CoM 3 LangMem 4 10.6600.6600.6810.660 20.6520.6520.6700.652 30.6530.6530.6710.653 40.6530.6530.6710.653 50.6550.6550.6730.655 60.6550.6550.6760.655 70.6610.6610.6810.661 80.6580.6580.6770.658 90.6670.6670.6880.667 100.6640.6640.6870.664 110.6580.6580.6750.658 120.6610.6610.6820.661 130.6620.6620.6810.662 140.6550.6550.6730.655 150.6510.6510.6690.651 160.6630.6630.6850.663 170.6590.6590.6800.659 180.6540.6540.6750.654 190.6650.6650.6820.665 200.6650.6650.6840.665 210.6560.6560.6700.656 220.6660.6660.6820.666 230.6640.6640.6850.664 240.6590.6590.6790.659 250.6530.6530.6710.653 260.6550.6550.6720.655 270.6580.6580.6770.658 280.6540.6540.6700.654 290.6580.6580.6740.658 300.6540.6540.6710.654 Table 8: All 30 consecutive read-only ED probes af- ter write-back. Each value averages the same 190 source questions in an architecture; these values un- derlie WEP 30 . 1 A-MEM (Xu et al., 2025); 2 Mem0 (Chhikara et al., 2025); 3 CoM (Xu et al., 2026b); 4 LangMem (LangChain Team, 2025). is near their Baseline ED, yet WEP falls sharply. The complete 30-query curve therefore carries in- formation that a single final search cannot provide. In A-MEM and Mem0, the answer check can use a concise verified span, producing larger accuracy gains. The side-by-side model blocks show that this split appears with both generators, while the mag- nitude of each gain remains architecture-dependent. MemoryBaseline FPAGain [95% CI] A-MEM0.818 1.0000.182 [0.145, 0.218] Mem00.821 1.0000.179 [0.142, 0.216] CoM0.729 0.8260.097 [0.066, 0.132] LangMem0.758 0.8390.082 [0.050, 0.116] Table 9: Kimi K2.6 accuracy across memory archi- tectures (190 source questions). Gain is FPA minus Baseline; brackets show 95% bootstrap intervals. Archi- tectures: A-MEM (Xu et al., 2025), Mem0 (Chhikara et al., 2025), CoM (Xu et al., 2026b), and Lang- Mem (LangChain Team, 2025). C Components and cross-model gains Table 11 retains every DeepSeek component con- figuration, while Table 9 reports the paired Kimi Baseline-to-FPA change by architecture. Both use the same source questions as the main method ta- ble; the Kimi intervals resample 190 question-level differences. The first two component rows isolate differ- ent tasks. Family control limits multiplicity, and provenance removes old-answer eligibility, but nei- ther by itself gives the generator the replacement fact. Adding authoritative evidence to Family with Provenance changes accuracy from 0.637 to 0.912; the final answer check adds a smaller further gain. Table 9 then tests whether the complete pipeline’s accuracy effect survives a change of generator rather than assigning model transfer to any single component. D DeepSeek task and source tests Task transfer and source quality test distinct condi- tions for the gate. Table 12 asks whether historical- answer adoption is present in a new task; Table 13 asks whether a usable correction accompanies the filtering rule. Every source-condition row uses the same 760 paired states. “Valid history retained” measures storage retention and equals 1.0 through- out these FPA runs. Stored history can remain intact even when a record is not served to the gen- erator. The shopping and medical tasks begin with sub- stantial old-answer adoption, and both gain ac- curacy as that rate falls. Travel begins with no old-answer adoption, so reducing access to histori- cal planning context does not address a measured Echo; its accuracy decreases. The source test iden- tifies a second boundary. Family-ID noise leaves some provenance uncertain but still permits a re- liable correction to be used. When the correction itself is missing or noisy, the gate can withhold old context without supplying a valid replacement. The retained-history column confirms that this is a serving failure, not deletion of stored records. E Reproducibility details TheA–HCSVsandjudgedan- swersremainintheexperiment’s results_memory_echo_ah_completeand associated result directories, separate from this minimal Overleaf package. Accuracy uses correct- ness scores of at least 0.75; old-answer rate uses a 12 DeepSeek V4 Flash MethodAcc.OldEDWEPDCE A-MEM 1 Baseline0.8340.0970.6610.4240.097 Budget0.9450.0180.4740.1110.013 CRAG-M0.9320.0210.7490.1090.016 PairLR0.9340.0210.7490.2330.016 FPA1.0000.0000.3760.0030.000 Mem0 2 Baseline0.8240.1050.6670.4270.105 Budget0.9450.0210.5010.1160.017 CRAG-M0.9320.0210.7580.1130.016 PairLR0.9370.0210.7580.2340.016 FPA1.0000.0000.4030.0030.000 CoM 3 Baseline0.8130.1160.6610.4340.116 Budget0.7680.1030.6610.4230.100 CRAG-M0.7710.0970.6610.3880.095 PairLR0.7840.0970.6310.3940.095 FPA0.8320.0630.6660.1380.066 LangMem 4 Baseline0.8290.1050.6610.4240.105 Budget0.7870.0950.6610.4070.093 CRAG-M0.7870.0950.6610.3720.093 PairLR0.7870.0920.6310.3770.089 FPA0.8500.0500.6660.0870.053 Kimi K2.6 MethodAcc.OldEDWEPDCE A-MEM 1 Baseline0.8180.1000.6610.4240.100 Budget0.9550.0210.4740.1110.017 CRAG-M0.9450.0210.7490.1090.016 PairLR0.9420.0240.7490.2330.019 FPA1.0000.0000.3760.0030.000 Mem0 2 Baseline0.8210.1000.6670.4270.101 Budget0.9470.0210.5010.1160.018 CRAG-M0.9390.0290.7580.1130.026 PairLR0.9420.0260.7580.2340.022 FPA1.0000.0000.4030.0030.000 CoM 3 Baseline0.7290.1210.6610.4340.121 Budget0.7340.1340.6610.4230.139 CRAG-M0.7370.1080.6610.3880.112 PairLR0.7420.1110.6310.3940.105 FPA0.8260.0760.6660.1380.081 LangMem 4 Baseline0.7580.1110.6610.4240.115 Budget0.7530.1240.6610.4070.118 CRAG-M0.7470.1000.6610.3720.104 PairLR0.7550.1130.6310.3770.114 FPA0.8390.0740.6660.0870.073 Table 10: Main-method results by architecture: DeepSeek V4 Flash (left) and Kimi K2.6 (right). Each architecture– method cell contains 380 reinforced-history states. 1 A-MEM (Xu et al., 2025); 2 Mem0 (Chhikara et al., 2025); 3 CoM (Xu et al., 2026b); 4 LangMem (LangChain Team, 2025); CRAG-M (Yan et al., 2024); PairLR (Burges et al., 2005). Retrieved-memory measurements are identical across the two generators. Baseline Budget CRAG-M PairLR FPA 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (higher is better) Baseline Budget CRAG-M PairLR FPA 0.00 0.02 0.04 0.06 0.08 0.10 0.12 DCE (lower is better) DeepSeek V4 FlashKimi K2.6 Figure 8: Table D visualized by generator. Vertical bars compare accuracy and DCE for five serving methods on 1,520 paired answer states per model. DeepSeek and Kimi are shown separately; exact values appear in Table 2. separate binary contamination judgment, and DCE takes the positive Echo-minus-gold alignment difference. All 30 delay probes are read-only. Main-method means cover two reinforced histories per question and architecture, balanced across datasets and architectures after reserving ten LoCoMo development questions. The 25 selected case traces and all 30 aggregate delay means are reproduced in the appendix. For a candidate memory recordm, family repre- 13 ConfigurationAcc.Old DCE Baseline0.825 0.106 0.106 Family only0.651 0.095 0.097 Provenance only0.820 0.010 0.009 Family w/ Provenance0.637 0.012 0.011 w/ Authority0.912 0.031 0.032 w/ Answer check (FPA)0.920 0.028 0.030 Table 11: Experiment E: DeepSeek FPA components; 1,520 paired states per configuration. Family and Prove- nance are tested separately before the components are added sequentially. This is an ablation of the FPA method; the compared retrieval and ranker families in Table 2 follow Yan et al. (2024); Burges et al. (2005). TaskBAccFAccBOldFOld∆Old Shopping 1 0.6100.8700.3050.020-0.285 Medical QA 2 0.5850.8500.3950.040-0.355 Travel 3 0.7850.6900.0000.0000.000 Table 12: DeepSeek task transfer (200 answer states per method and task). B and F denote Baseline and FPA; Acc is accuracy, Old is old-answer rate, and∆Old is FPA minus Baseline. 1 WebShop (Yao et al., 2022); 2 MedQA (Jin et al., 2021); 3 TravelPlanner (Xie et al., 2024). Source conditionBase acc. FPA acc. Retained Complete0.6040.6291.000 Family ID noise0.5990.6301.000 Correction missing/noisy0.3780.2121.000 Old fact nonconflicting0.6640.6751.000 Table 13: DeepSeek source-quality test; 760 answer states per method and condition. FPA retains valid his- torical records in storage. Source questions: LoCoMo (Maharana et al., 2024), LongMemEval (Wu et al., 2025), HotpotQA (Yang et al., 2018), BEAM (Tavakoli et al., 2026). sentative selection uses S(m) = s ret (m) + 0.35c(m) + 0.80r(m) − 0.18 log(1 + v(m) + n(m)) − 4I stale (m), (5) Symbols.mis a candidate memory record and S(m)its score for selecting one representative from its source-ancestry family.s ret (m)is the original retrieval score,c(m)the record’s confi- dence, andr(m)its source reliability.v(m)counts writes and n(m) counts retrieval exposures. log is the natural logarithm, so the penalty for repeated writes and exposures grows sublinearly; the added 1makes the expression defined when both counts are zero.I stale (m)equals1for a superseded or suppressed record and0otherwise. The numeric coefficients are fixed scoring weights, not probabili- ties. The highest-scoring eligible member takes the family’s single serving position; its other records do not become independent evidence. F Interpretation of the Echo lifecycle Memory Echo arises when one source claim be- comes several retrievable records. These deriva- tives can crowd out a stored correction through 30 read-only probes. ED measures retrieval oc- cupancy, RRI correction resistance, WEP delayed exposure, and DCE influence on the final answer. Occupancy need not cause an incorrect answer; suppressing history without replacement evidence can also fail. FPA caps family occupancy, gates superseded claims, and checks answers against authoritative evidence. Source-quality tests show the need for reliable replacement evidence; fixed-candidate re- play isolates serving effects from native memory evolution. G Comparator protocols and information access Budget feeds occupancy above a Seed-only limit into a bounded feedback controller: h t = min(3, 0.85h t−1 + 1.25p t ).(6) Symbols.fis the memory architecture,sthe lifecy- cle stage, andb f,s their Seed-only ED budget (the 95th percentile). At turnt,E t is old-family ED and p t = max(0,E t − b f,s )is its excess above bud- get.h t−1 andh t are previous and current feedback strengths, initialized at zero.0.85retains earlier pressure,1.25weights new excess, andmin(3,·) caps the state at three. The state penalizes exposed or stale candidates and favors verified corrections; its signal requires labeled ED. CRAG-M grades each retrieved candidate for reranking: Q(m) = 0.45r + 0.35v + 0.15c− 0.45u.(7) Symbols.mis a retrieved candidate andQ(m) its quality score, not a probability.r ∈ [−1, 1] is clipped retrieval relevance;vis the verified- external-source indicator;c∈ [0, 1]is confidence; uis the conflict-or-stale indicator. Both indica- tors are binary. The fixed positive coefficients re- wardr,v,c, while the negative coefficient penal- izesu. Low-quality candidates are demoted; ab- sent reliable correction, a corrective-evidence flag is recorded. This adaptation of CRAG (Yan et al., 2024) omits its learned evaluator and web search. PairLR learns a preference between a corrected candidate and an Echo candidate using retrieval, 14 confidence, activation, write/exposure, provenance, and status features: P (m + ≻ m − ) = σ(θ ⊤ ∆φ).(8) Symbols.m + is the corrected candidate andm − the Echo candidate;≻means “should rank above.” φ(m)is a candidate’s feature vector, and∆φ = φ(m + )− φ(m − )is the corrected-minus-Echo fea- ture difference.θis the learned weight vector, θ ⊤ ∆φtheir dot product, andσ(z) = 1/(1 + e −z ) the logistic function, withzits input andethe ex- ponential base.Pis the predicted probability of preferring the corrected candidate. Five folds are split by source question, so each held-out question is scored without its own pair in training. PairLR reranks after correction without changing language- model weights; the comparison concerns candidate serving, not policy optimization. MethodInformation used and serving operation FPANo Echo label for ranking; source-family IDs supplied. One position per family; verified correction supersedes history; authority guides and checks the answer. BudgetLabeled ED sets the pressure and Seed-only budget. Exposure, conflict, and delay signals penalize excess occupancy. CRAG-M No Echo label in its quality score. Uses verified provenance, confidence, and stale status to rerank candidates and flag an absent correction. PairLRLabeled pairs for training only. Uses retrieval and provenance features for five-fold, question-held-out pairwise reranking. Table 14: Comparator information in paired candidate- pool replay. CRAG-M adapts Yan et al. (2024); PairLR follows Burges et al. (2005). H Supplementary visual comparisons Figure 9 shows the pooled DeepSeek levels be- hind the method contrasts.Figure 10 shows method-paired effects across the same question– architecture–history states as Table 2. ED, WEP, and DCE in the lollipop chart reproduce Table 2; RRI averages storage-side resistance from the matched lifecycle state and each method’s serving- side resistance. 00.51 Risk FPA Budget Baseline PairLR CRAG-M 0.528 0.574 0.662 0.692 0.707 (a) ED ↓ 00.51 Risk FPA CRAG-M Budget PairLR Baseline 0.394 0.544 0.583 0.635 0.653 (b) RRI ↓ 00.51 Risk FPA CRAG-M Budget PairLR Baseline 0.057 0.245 0.264 0.309 0.427 (c) WEP ↓ 00.51 Risk FPA PairLR CRAG-M Budget Baseline 0.030 0.054 0.055 0.056 0.106 (d) DCE ↓ Baseline Budget CRAG-M PairLR FPA Figure 9: Current DeepSeek method means for the four Echo measures, recomputed from the final 1,520-state Table D cohort. Lower is better in every panel; values are printed beside the marks. −0.4−0.3−0.2−0.10.0 FPA − comparator Budget CRAG-M PairLR (a) ED ↓ −0.4−0.3−0.2−0.10.0 FPA − comparator (b) RRI ↓ −0.4−0.3−0.2−0.10.0 FPA − comparator Budget CRAG-M PairLR (c) WEP ↓ −0.4−0.3−0.2−0.10.0 FPA − comparator (d) DCE ↓ Figure 10: Paired DeepSeek effects on 190 questions, four architectures, and two write-back histories. Bars show mean FPA-minus-comparator effects; dark marks show 95% source-question bootstrap intervals (10,000 resamples). Negative values favor FPA. AbilityQEDRRIWEPDCE LoCoMo Adversarial20.5580.7450.5630.000 Multi-hop100.6000.7600.5810.075 Open-domain90.5910.7680.5600.139 Single-hop90.6460.8000.5680.472 Temporal100.5930.7650.5750.225 LongMemEval Update100.6670.8030.6080.300 Multi-session100.8060.8820.6520.625 Assistant recall100.6780.7950.6620.400 User recall100.7570.8570.7540.600 Temporal100.7240.8190.6750.300 Table 15: Ability-level means in Figure 5. Q counts source questions; each contributes four framework states. Sources: LoCoMo (Maharana et al., 2024) and LongMemEval (Wu et al., 2025). 15 Legacy taskMethodPost EDRRIWEPDCE ShoppingBaseline0.5864 0.7951 0.5861 0.8188 FPA0.0042 0.4025 0.0043 0.1683 Medical QA Baseline0.3514 0.7060 0.3422 0.6658 FPA0.0157 0.4002 0.0137 0.3268 TravelBaseline0.5436 0.7641 0.5401 0.8538 FPA0.0000 0.3820 0.0000 0.1436 Table 16: Values for the earlier extractive pilot in Fig- ure 7, not the current Table G cohort. Tasks use Web- Shop (Yao et al., 2022), MedQA (Jin et al., 2021), and TravelPlanner (Xie et al., 2024). I Refreshed five-stage case catalogue The previous appendix selected five illustrative cases for each lifecycle stage. We preserve that organization but recompute all retrieval counts, metrics, and recorded answers from the latest 190- question lifecycle and judged-answer files. The selected questions come from LoCoMo (Maharana et al., 2024), LongMemEval (Wu et al., 2025), Hot- potQA (Yang et al., 2018), and BEAM (Tavakoli et al., 2026). The cards show the actual question and candidate answer strings, the retrieved records that made the historical claim visible, and the down- stream generated answer. They are mechanism examples rather than prevalence estimates. Stage probes are retrieval-only; their associated Final De- cision answer is a later observation from the same trajectory. Data refresh. The 25 case slots and five-stage grouping follow the previous appendix, but every question, retrieval, metric, and answer below is taken from the current 190-question retrieve_writebacklifecycle and DeepSeek V4 Flash/GPT-5.6 judged-answer records. These are illustrative selections, not an incidence estimate. Each stage probe reads the same trajectory; only the Final Decision stage generates the recorded answer. RRI and WEP are trajectory-level mea- sures, not additional per-probe judgments. Four former LoCoMo source IDs fell within the ten de- velopment questions excluded from this run. Their case slots use same-dataset, same-ability held-out replacements:multi-00025→multi-00045, multi-00026→multi-00046, multi-00027→multi-00047,and multi-00029→multi-00049.No older case metric is mixed into the new results. I.1 Encoding Case 1.1: multi-00068| LongMemEval| Mem0 Ability and question. single-session-assistant. I’m planning to revisit Orlando. I was wondering if you could remind me of that unique dessert shop with the giant milkshakes we talked about last time? Verified answer / historical seed. The Sugar Fac- tory at Icon Park. / I mentioned biometric authenti- cation or one-time passwords (OTP) as examples of two-factor authentication methods. Stage probe. Using stored memory, answer: I’m planning to revisit Orlando. I was wondering if you could remind me of that unique dessert shop with the giant milkshakes we talked about last time? Observed retrieval. Top-4: 1 Echo-family and 0 correction record(s); stage ED 0.390. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’I’m planning to revisit Orlando. I was wondering if you could remind me of that unique dessert shop with the giant milk- shakes we talked about last time?’ as: I mentioned biometric authentication or one-time passwords (OTP) as examples of two-factor authentication methods.; #2 NOISE: Unrelated background record 2 for multi-00068; it contains no answer to the tar- get question.; #3 NOISE: Unrelated background record 1 for multi-00068; it contains no answer to the target question. Lifecycle measures. Pre/post ED 0.976/0.541; re- trieval RRI 0.584; WEP 30 0.541; ED at delay 1/30 0.541/0.541. Downstream answer (same trajectory). The Sugar Factory at Icon Park. Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The seed’s initial retrieval rank and ED show how much space it receives before practice. Case 1.2: multi-00036| LoCoMo| Mem0 Ability and question. multi-hop. Would Caroline be considered religious? Verified answer / historical seed. Somewhat, but not extremely religious / Pomodoro technique Stage probe.Using stored memory, answer: Would Caroline be considered religious? Observed retrieval. Top-4: 1 Echo-family and 0 correction record(s); stage ED 0.390. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’Would Caroline be con- sidered religious?’ as: Pomodoro technique; #2 NOISE: Unrelated background record 1 for multi- 00036; it contains no answer to the target ques- tion.; #3 NOISE: Unrelated background record 0 for multi-00036; it contains no answer to the target question. 16 Lifecycle measures. Pre/post ED 0.908/0.569; re- trieval RRI 0.627; WEP 30 0.569; ED at delay 1/30 0.569/0.569. Downstream answer (same trajectory). Some- what, but not extremely religious. Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The seed’s initial retrieval rank and ED show how much space it receives before practice. Case 1.3: multi-00012| LoCoMo| Mem0 Ability and question. multi-hop. Would Caroline likely have Dr. Seuss books on her bookshelf? Verified answer / historical seed. Yes, since she collects classic children’s books / Middle-class or wealthy Stage probe.Using stored memory, answer: Would Caroline likely have Dr. Seuss books on her bookshelf? Observed retrieval. Top-4: 1 Echo-family and 0 correction record(s); stage ED 0.390. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’Would Caroline likely have Dr. Seuss books on her bookshelf?’ as: Middle-class or wealthy; #2 NOISE: Unrelated background record 2 for multi-00012; it contains no answer to the target question.; #3 NOISE: Un- related background record 1 for multi-00012; it contains no answer to the target question. Lifecycle measures. Pre/post ED 0.930/0.596; re- trieval RRI 0.639; WEP 30 0.596; ED at delay 1/30 0.596/0.596. Downstream answer (same trajectory). Yes, since she collects classic children’s books. Correct- ness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The seed’s initial retrieval rank and ED show how much space it receives before practice. Case 1.4: multi-00042| LoCoMo| Mem0 Ability and question. single-hop. What LGBTQ+ events has Caroline participated in? Verified answer / historical seed. Pride parade, school speech, support group / A doll, a film cam- era Stage probe. Using stored memory, answer: What LGBTQ+ events has Caroline participated in? Observed retrieval. Top-4: 1 Echo-family and 0 correction record(s); stage ED 0.390. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’What LGBTQ+ events has Caroline participated in?’ as: A doll, a film camera; #2 NOISE: Unrelated background record 2 for multi-00042; it contains no answer to the tar- get question.; #3 NOISE: Unrelated background record 1 for multi-00042; it contains no answer to the target question. Lifecycle measures. Pre/post ED 0.919/0.564; re- trieval RRI 0.619; WEP 30 0.583; ED at delay 1/30 0.564/0.564. Downstream answer (same trajectory). Pride parade, school speech, support group. Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The seed’s initial retrieval rank and ED show how much space it receives before practice. Case 1.5: multi-00014| LoCoMo| Mem0 Ability and question. single-hop. Where did Car- oline move from 4 years ago? Verified answer / historical seed. Sweden / ab- stract art Stage probe.Using stored memory, answer: Where did Caroline move from 4 years ago? Observed retrieval. Top-4: 1 Echo-family and 0 correction record(s); stage ED 0.390. Top retrieved records. #1 ECHO: Historical memory answers the question ’Where did Caro- line move from 4 years ago?’ as: abstract art; #2 NOISE: Unrelated background record 2 for multi- 00014; it contains no answer to the target ques- tion.; #3 NOISE: Unrelated background record 1 for multi-00014; it contains no answer to the target question. Lifecycle measures. Pre/post ED 0.887/0.548; re- trieval RRI 0.618; WEP 30 0.587; ED at delay 1/30 0.548/0.548. Downstream answer (same trajectory). Sweden Correctness 1.000; DCE 0.000; old-answer con- tamination: no. What this stage tests. The seed’s initial retrieval rank and ED show how much space it receives before practice. I.2 Retrieval Practice Case 2.1: multi-00194| BEAM| Mem0 Ability and question. abstention. What was the atmosphere like during the February 20 book club discussion on ’The Poppy War’ hosted by Kelly and I? Verified answer / historical seed. Based on the provided chat, there is no information related to 17 the atmosphere during the February 20 book club discussion on ’The Poppy War’. / Based on the provided chat, there is no information related to the topics or skills covered in the advanced storytelling workshop starting September 15. Stage probe. Retrieve the historical memory and answer: What was the atmosphere like during the February 20 book club discussion on ’The Poppy War’ hosted by Kelly and I? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.982. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’What was the atmosphere like during the February 20 book club discussion on ’The Poppy War’ hosted by Kelly and I?’ as: Based on the provided chat, there is no informa- tion related to the topics or skills covered in the advanced storytelling workshop starting September 15. Reconsolidation count 5.; #2 ECHO: Histori- cal memory answers the question ’What was the atmosphere like during the February 20 book club discussion on ’The Poppy War’ hosted by Kelly and I?’ as: Based on the provided chat, there is no information related to the topics or skills cov- ered in the advanced storytelling workshop starting September 15. Retrieval-derived trace 5 repeats the historical answer for future use.; #3 ECHO: Histor- ical memory answers the question ’What was the atmosphere like during the February 20 book club discussion on ’The Poppy War’ hosted by Kelly and I?’ as: Based on the provided chat, there is no information related to the topics or skills cov- ered in the advanced storytelling workshop starting September 15. Retrieval-derived trace 1 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.982/0.605; re- trieval RRI 0.629; WEP 30 0.605; ED at delay 1/30 0.605/0.605. Downstream answer (same trajectory). There is no information available about the atmosphere dur- ing that book club discussion. Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. Six write-backs have oc- curred; compare this stage ED with pre-correction ED to locate amplification. Case 2.2: multi-00186| BEAM| Mem0 Ability and question. temporal-reasoning. How many days did it take me to finish reading the tril- ogy after I downloaded it? Verified answer / historical seed. It took 12 days to finish reading the trilogy after downloading it on December 7. / I started setting clear work-free Sundays 14 days after my weekend getaway with David on April 20-21, beginning on May 5. Stage probe. Retrieve the historical memory and answer: How many days did it take me to finish reading the trilogy after I downloaded it? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.972. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’How many days did it take me to finish reading the trilogy after I down- loaded it?’ as: I started setting clear work-free Sun- days 14 days after my weekend getaway with David on April 20-21, beginning on May 5. Retrieval- derived trace 3 repeats the historical answer for future use.; #2 ECHO: Historical memory answers the question ’How many days did it take me to fin- ish reading the trilogy after I downloaded it?’ as: I started setting clear work-free Sundays 14 days after my weekend getaway with David on April 20- 21, beginning on May 5. Retrieval-derived trace 2 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’How many days did it take me to finish reading the trilogy after I downloaded it?’ as: I started setting clear work-free Sundays 14 days after my week- end getaway with David on April 20-21, beginning on May 5. Retrieval-derived trace 4 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.976/0.642; re- trieval RRI 0.656; WEP 30 0.648; ED at delay 1/30 0.642/0.642. Downstream answer (same trajectory). 12 days Correctness 1.000; DCE 0.000; old-answer con- tamination: no. What this stage tests. Six write-backs have oc- curred; compare this stage ED with pre-correction ED to locate amplification. Case 2.3: multi-00068| LongMemEval| Mem0 Ability and question. single-session-assistant. I’m planning to revisit Orlando. I was wondering if you could remind me of that unique dessert shop with the giant milkshakes we talked about last time? Verified answer / historical seed. The Sugar Fac- tory at Icon Park. / I mentioned biometric authenti- cation or one-time passwords (OTP) as examples of two-factor authentication methods. Stage probe. Retrieve the historical memory and answer: I’m planning to revisit Orlando. I was wondering if you could remind me of that unique 18 dessert shop with the giant milkshakes we talked about last time? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.969. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’I’m planning to revisit Orlando. I was wondering if you could remind me of that unique dessert shop with the giant milk- shakes we talked about last time?’ as: I mentioned biometric authentication or one-time passwords (OTP) as examples of two-factor authentication methods. Retrieval-derived trace 3 repeats the his- torical answer for future use.; #2 ECHO: Histor- ical memory answers the question ’I’m planning to revisit Orlando. I was wondering if you could remind me of that unique dessert shop with the giant milkshakes we talked about last time?’ as: I mentioned biometric authentication or one-time passwords (OTP) as examples of two-factor authen- tication methods. Retrieval-derived trace 4 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’I’m plan- ning to revisit Orlando. I was wondering if you could remind me of that unique dessert shop with the giant milkshakes we talked about last time?’ as: I mentioned biometric authentication or one-time passwords (OTP) as examples of two-factor authen- tication methods. Retrieval-derived trace 5 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.976/0.541; re- trieval RRI 0.584; WEP 30 0.541; ED at delay 1/30 0.541/0.541. Downstream answer (same trajectory). The Sugar Factory at Icon Park. Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. Six write-backs have oc- curred; compare this stage ED with pre-correction ED to locate amplification. Case 2.4: multi-00181| BEAM| Mem0 Ability and question. temporal-reasoning. How many days had I been journaling when I noted my 40% improvement in decision clarity? Verified answer / historical seed. I had been jour- naling daily for 58 days when I noted my 40% im- provement in decision clarity on May 31. / There are about 4 months between April 15, 2024, when I planned to reach my daily walking goal, and Au- gust 22, 2024, the date of the festival I’m preparing m... Stage probe. Retrieve the historical memory and answer: How many days had I been journaling when I noted my 40% improvement in decision clarity? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.963. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’How many days had I been journaling when I noted my 40% improve- ment in decision clarity?’ as: There are about 4 months between April 15, 2024, when I planned to reach my daily walking goal, and August 22, 2024, the date of the festival I’m preparing m... Retrieval-derived trace 5 repeats the historical an- swer for future use.; #2 ECHO: Historical memory answers the question ’How many days had I been journaling when I noted my 40% improvement in decision clarity?’ as: There are about 4 months be- tween April 15, 2024, when I planned to reach my daily walking goal, and August 22, 2024, the date of the festival I’m preparing m... Retrieval-derived trace 3 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the ques- tion ’How many days had I been journaling when I noted my 40% improvement in decision clarity?’ as: There are about 4 months between April 15, 2024, when I planned to reach my daily walking goal, and August 22, 2024, the date of the festival I’m preparing m... Retrieval-derived trace 2 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.967/0.561; re- trieval RRI 0.601; WEP 30 0.561; ED at delay 1/30 0.561/0.561. Downstream answer (same trajectory). 58 days Correctness 1.000; DCE 0.000; old-answer con- tamination: no. What this stage tests. Six write-backs have oc- curred; compare this stage ED with pre-correction ED to locate amplification. Case 2.5: multi-00087| LongMemEval| Mem0 Ability and question. multi-session. How many days did I take social media breaks in total? Verified answer / historical seed. 17 days / 12 Stage probe. Retrieve the historical memory and answer: How many days did I take social media breaks in total? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.886. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’How many days did I take social media breaks in total?’ as: 12 Retrieval- derived trace 3 repeats the historical answer for 19 future use.; #2 ECHO: Historical memory answers the question ’How many days did I take social me- dia breaks in total?’ as: 12 Retrieval-derived trace 4 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’How many days did I take social media breaks in total?’ as: 12 Retrieval-derived trace 1 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.902/0.587; re- trieval RRI 0.642; WEP 30 0.573; ED at delay 1/30 0.573/0.573. Downstream answer (same trajectory). 12 Cor- rectness 0.000; DCE 1.000; old-answer contamina- tion: yes. What this stage tests. Six write-backs have oc- curred; compare this stage ED with pre-correction ED to locate amplification. I.3 Updating / Reconsolidation Case 3.1: multi-00046| LoCoMo| A-MEM Ability and question. single-hop. What events has Caroline participated in to help children? Verified answer / historical seed. Mentoring pro- gram, school speech / Poetry, creative writing Stage probe. Use the verified current memory to answer: What events has Caroline participated in to help children? Observed retrieval. Top-5: 4 Echo-family and 1 correction record(s); stage ED 0.549. Top retrieved records. #1 CORRECTION: Veri- fied evidence corrects the answer to ’What events has Caroline participated in to help children?’ as: Mentoring program, school speech. Supporting evidence: Evidence turns: D9:2, D3:3; #2 ECHO: The historical answer for ’What events has Car- oline participated in to help children?’ is super- seded and must not control the current answer.; #3 ECHO: Historical memory answers the question ’What events has Caroline participated in to help children?’ as: Poetry, creative writing Retrieval- derived trace 6 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.901/0.549; re- trieval RRI 0.614; WEP 30 0.549; ED at delay 1/30 0.549/0.549. Downstream answer (same trajectory). Mentor- ing program, school speech. Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The correction has ar- rived; simultaneous old and corrected records re- veal whether revision displaces the Echo. Case 3.2: multi-00046| LoCoMo| Mem0 Ability and question. single-hop. What events has Caroline participated in to help children? Verified answer / historical seed. Mentoring pro- gram, school speech / Poetry, creative writing Stage probe. Use the verified current memory to answer: What events has Caroline participated in to help children? Observed retrieval. Top-5: 4 Echo-family and 1 correction record(s); stage ED 0.549. Top retrieved records. #1 CORRECTION: Veri- fied evidence corrects the answer to ’What events has Caroline participated in to help children?’ as: Mentoring program, school speech. Supporting evidence: Evidence turns: D9:2, D3:3; #2 ECHO: The historical answer for ’What events has Car- oline participated in to help children?’ is super- seded and must not control the current answer.; #3 ECHO: Historical memory answers the question ’What events has Caroline participated in to help children?’ as: Poetry, creative writing Retrieval- derived trace 6 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.901/0.549; re- trieval RRI 0.614; WEP 30 0.549; ED at delay 1/30 0.549/0.549. Downstream answer (same trajectory). Mentor- ing program, school speech. Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The correction has ar- rived; simultaneous old and corrected records re- veal whether revision displaces the Echo. Case 3.3: multi-00049| LoCoMo| A-MEM Ability and question. open-domain. What was grandma’s gift to Caroline? Verified answer / historical seed. necklace / 7 years Stage probe. Use the verified current memory to answer: What was grandma’s gift to Caroline? Observed retrieval. Top-5: 4 Echo-family and 1 correction record(s); stage ED 0.568. Top retrieved records. #1 CORRECTION: Ver- ified evidence corrects the answer to ’What was grandma’s gift to Caroline?’ as: necklace. Sup- porting evidence: Evidence turns: D4:3; #2 ECHO: Historical memory answers the question ’What was grandma’s gift to Caroline?’ as: 7 years Retrieval- derived trace 6 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’What was grandma’s gift to Caro- 20 line?’ as: 7 years Retrieval-derived trace 2 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.883/0.568; re- trieval RRI 0.635; WEP 30 0.566; ED at delay 1/30 0.566/0.566. Downstream answer (same trajectory). neck- lace Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The correction has ar- rived; simultaneous old and corrected records re- veal whether revision displaces the Echo. Case 3.4: multi-00049| LoCoMo| Mem0 Ability and question. open-domain. What was grandma’s gift to Caroline? Verified answer / historical seed. necklace / 7 years Stage probe. Use the verified current memory to answer: What was grandma’s gift to Caroline? Observed retrieval. Top-5: 4 Echo-family and 1 correction record(s); stage ED 0.568. Top retrieved records. #1 CORRECTION: Ver- ified evidence corrects the answer to ’What was grandma’s gift to Caroline?’ as: necklace. Sup- porting evidence: Evidence turns: D4:3; #2 ECHO: Historical memory answers the question ’What was grandma’s gift to Caroline?’ as: 7 years Retrieval- derived trace 6 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’What was grandma’s gift to Caro- line?’ as: 7 years Retrieval-derived trace 2 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.883/0.568; re- trieval RRI 0.635; WEP 30 0.566; ED at delay 1/30 0.566/0.566. Downstream answer (same trajectory). neck- lace Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The correction has ar- rived; simultaneous old and corrected records re- veal whether revision displaces the Echo. Case 3.5: multi-00047| LoCoMo| Mem0 Ability and question. temporal-reasoning. When did Caroline have a picnic? Verified answer / historical seed. The week before 6 July 2023 / 24 April, 2023 Stage probe. Use the verified current memory to answer: When did Caroline have a picnic? Observed retrieval. Top-5: 4 Echo-family and 1 correction record(s); stage ED 0.583. Top retrieved records. #1 CORRECTION: Ver- ified evidence corrects the answer to ’When did Caroline have a picnic?’ as: The week before 6 July 2023. Supporting evidence: Evidence turns: D6:11; #2 ECHO: Historical memory answers the question ’When did Caroline have a picnic?’ as: 24 April, 2023 Retrieval-derived trace 6 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’When did Caroline have a picnic?’ as: 24 April, 2023 Retrieval-derived trace 5 repeats the historical an- swer for future use. Lifecycle measures. Pre/post ED 0.905/0.583; re- trieval RRI 0.638; WEP 30 0.571; ED at delay 1/30 0.571/0.571. Downstream answer (same trajectory). 24 April, 2023 Correctness 0.000; DCE 1.000; old-answer contamination: yes. What this stage tests. The correction has ar- rived; simultaneous old and corrected records re- veal whether revision displaces the Echo. I.4 Delayed Retrieval Case 4.1: multi-00079| LongMemEval| A-MEM Ability and question. single-session-user. What was my last name before I changed it? Verified answer / historical seed. Johnson / 20 Stage probe. After 30 later read-only sessions, answer from current memory: What was my last name before I changed it? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.869. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’What was my last name before I changed it?’ as: 20 Retrieval-derived trace 5 repeats the historical answer for future use.; #2 ECHO: Historical memory answers the question ’What was my last name before I changed it?’ as: 20 Retrieval-derived trace 4 repeats the historical answer for future use.; #3 ECHO: Historical mem- ory answers the question ’What was my last name before I changed it?’ as: 20 Retrieval-derived trace 6 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.868/0.869; re- trieval RRI 1.000; WEP 30 0.869; ED at delay 1/30 0.868/0.869. Downstream answer (same trajectory). 20 Cor- rectness 0.000; DCE 1.000; old-answer contamina- tion: yes. 21 What this stage tests. No further write occurs; the first-to-thirtieth ED and WEP distinguish transient from persistent exposure. Case 4.2: multi-00079| LongMemEval| Mem0 Ability and question. single-session-user. What was my last name before I changed it? Verified answer / historical seed. Johnson / 20 Stage probe. After 30 later read-only sessions, answer from current memory: What was my last name before I changed it? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.869. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’What was my last name before I changed it?’ as: 20 Retrieval-derived trace 5 repeats the historical answer for future use.; #2 ECHO: Historical memory answers the question ’What was my last name before I changed it?’ as: 20 Retrieval-derived trace 4 repeats the historical answer for future use.; #3 ECHO: Historical mem- ory answers the question ’What was my last name before I changed it?’ as: 20 Retrieval-derived trace 6 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.868/0.869; re- trieval RRI 1.000; WEP 30 0.869; ED at delay 1/30 0.868/0.869. Downstream answer (same trajectory). 20 Cor- rectness 0.000; DCE 1.000; old-answer contamina- tion: yes. What this stage tests. No further write occurs; the first-to-thirtieth ED and WEP distinguish transient from persistent exposure. Case 4.3: multi-00187| BEAM| A-MEM Ability and question.information-extraction. Which option did I say I chose after trying both at the store? Verified answer / historical seed. You said you chose the Adidas Ultraboost over the Nike React Infinity Run after trying both on March 30 at Foot Locker. / You have been with Douglas for 3 years. Stage probe. After 30 later read-only sessions, answer from current memory: Which option did I say I chose after trying both at the store? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.865. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’Which option did I say I chose after trying both at the store?’ as: You have been with Douglas for 3 years. Retrieval-derived trace 6 repeats the historical answer for future use.; #2 ECHO: Historical memory answers the question ’Which option did I say I chose after trying both at the store?’ as: You have been with Douglas for 3 years. Retrieval-derived trace 2 repeats the histor- ical answer for future use.; #3 ECHO: Historical memory answers the question ’Which option did I say I chose after trying both at the store?’ as: You have been with Douglas for 3 years. Retrieval- derived trace 1 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.879/0.823; re- trieval RRI 0.968; WEP 30 0.851; ED at delay 1/30 0.864/0.865. Downstream answer (same trajectory). You have been with Douglas for 3 years. Correctness 0.000; DCE 1.000; old-answer contamination: yes. What this stage tests. No further write occurs; the first-to-thirtieth ED and WEP distinguish transient from persistent exposure. Case 4.4: multi-00187| BEAM| Mem0 Ability and question.information-extraction. Which option did I say I chose after trying both at the store? Verified answer / historical seed. You said you chose the Adidas Ultraboost over the Nike React Infinity Run after trying both on March 30 at Foot Locker. / You have been with Douglas for 3 years. Stage probe. After 30 later read-only sessions, answer from current memory: Which option did I say I chose after trying both at the store? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.865. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’Which option did I say I chose after trying both at the store?’ as: You have been with Douglas for 3 years. Retrieval-derived trace 6 repeats the historical answer for future use.; #2 ECHO: Historical memory answers the question ’Which option did I say I chose after trying both at the store?’ as: You have been with Douglas for 3 years. Retrieval-derived trace 2 repeats the histor- ical answer for future use.; #3 ECHO: Historical memory answers the question ’Which option did I say I chose after trying both at the store?’ as: You have been with Douglas for 3 years. Retrieval- derived trace 1 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.879/0.823; re- trieval RRI 0.968; WEP 30 0.851; ED at delay 1/30 0.864/0.865. 22 Downstream answer (same trajectory). You have been with Douglas for 3 years. Correctness 0.000; DCE 1.000; old-answer contamination: yes. What this stage tests. No further write occurs; the first-to-thirtieth ED and WEP distinguish transient from persistent exposure. Case 4.5: multi-00105| HotpotQA| Mem0 Ability and question. multi-hop-comparison. Are Local H and For Against both from the United States? Verified answer / historical seed. yes / MiMA Stage probe. After 30 later read-only sessions, answer from current memory: Are Local H and For Against both from the United States? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.868. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’Are Local H and For Against both from the United States?’ as: MiMA Retrieval-derived trace 2 repeats the historical an- swer for future use.; #2 ECHO: Historical memory answers the question ’Are Local H and For Against both from the United States?’ as: MiMA Retrieval- derived trace 1 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’Are Local H and For Against both from the United States?’ as: MiMA Retrieval- derived trace 5 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.881/0.868; re- trieval RRI 0.993; WEP 30 0.868; ED at delay 1/30 0.868/0.868. Downstream answer (same trajectory). Yes. Cor- rectness 1.000; DCE 0.000; old-answer contamina- tion: no. What this stage tests. No further write occurs; the first-to-thirtieth ED and WEP distinguish transient from persistent exposure. I.5 Final Decision Case 5.1: multi-00105| HotpotQA| A-MEM Ability and question. multi-hop-comparison. Are Local H and For Against both from the United States? Verified answer / historical seed. yes / MiMA Stage probe. Make the final factually valid deci- sion using memory: Are Local H and For Against both from the United States? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.868. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’Are Local H and For Against both from the United States?’ as: MiMA Retrieval-derived trace 2 repeats the historical an- swer for future use.; #2 ECHO: Historical memory answers the question ’Are Local H and For Against both from the United States?’ as: MiMA Retrieval- derived trace 1 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’Are Local H and For Against both from the United States?’ as: MiMA Retrieval- derived trace 6 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.881/0.868; re- trieval RRI 0.993; WEP 30 0.868; ED at delay 1/30 0.868/0.868. Downstream answer (same trajectory). Yes. Cor- rectness 1.000; DCE 0.000; old-answer contamina- tion: no. What this stage tests. The answer and DCE show whether the retrieved Echo crossed into the final response. Case 5.2: multi-00105| HotpotQA| Mem0 Ability and question. multi-hop-comparison. Are Local H and For Against both from the United States? Verified answer / historical seed. yes / MiMA Stage probe. Make the final factually valid deci- sion using memory: Are Local H and For Against both from the United States? Observed retrieval. Top-5: 5 Echo-family and 0 correction record(s); stage ED 0.868. Top retrieved records. #1 ECHO: Historical mem- ory answers the question ’Are Local H and For Against both from the United States?’ as: MiMA Retrieval-derived trace 2 repeats the historical an- swer for future use.; #2 ECHO: Historical memory answers the question ’Are Local H and For Against both from the United States?’ as: MiMA Retrieval- derived trace 1 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’Are Local H and For Against both from the United States?’ as: MiMA Retrieval- derived trace 6 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.881/0.868; re- trieval RRI 0.993; WEP 30 0.868; ED at delay 1/30 0.868/0.868. Downstream answer (same trajectory). Yes. Cor- rectness 1.000; DCE 0.000; old-answer contamina- tion: no. 23 What this stage tests. The answer and DCE show whether the retrieved Echo crossed into the final response. Case 5.3: multi-00046| LoCoMo| Mem0 Ability and question. single-hop. What events has Caroline participated in to help children? Verified answer / historical seed. Mentoring pro- gram, school speech / Poetry, creative writing Stage probe. Make the final factually valid de- cision using memory: What events has Caroline participated in to help children? Observed retrieval. Top-5: 4 Echo-family and 1 correction record(s); stage ED 0.582. Top retrieved records. #1 CORRECTION: Veri- fied evidence corrects the answer to ’What events has Caroline participated in to help children?’ as: Mentoring program, school speech. Supporting evi- dence: Evidence turns: D9:2, D3:3; #2 ECHO: His- torical memory answers the question ’What events has Caroline participated in to help children?’ as: Poetry, creative writing Retrieval-derived trace 5 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’What events has Caroline participated in to help children?’ as: Poetry, creative writing Retrieval- derived trace 6 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.901/0.549; re- trieval RRI 0.614; WEP 30 0.549; ED at delay 1/30 0.549/0.549. Downstream answer (same trajectory). Mentor- ing program, school speech. Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The answer and DCE show whether the retrieved Echo crossed into the final response. Case 5.4: multi-00046| LoCoMo| A-MEM Ability and question. single-hop. What events has Caroline participated in to help children? Verified answer / historical seed. Mentoring pro- gram, school speech / Poetry, creative writing Stage probe. Make the final factually valid de- cision using memory: What events has Caroline participated in to help children? Observed retrieval. Top-5: 4 Echo-family and 1 correction record(s); stage ED 0.582. Top retrieved records. #1 CORRECTION: Veri- fied evidence corrects the answer to ’What events has Caroline participated in to help children?’ as: Mentoring program, school speech. Supporting evi- dence: Evidence turns: D9:2, D3:3; #2 ECHO: His- torical memory answers the question ’What events has Caroline participated in to help children?’ as: Poetry, creative writing Retrieval-derived trace 5 repeats the historical answer for future use.; #3 ECHO: Historical memory answers the question ’What events has Caroline participated in to help children?’ as: Poetry, creative writing Retrieval- derived trace 6 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.901/0.549; re- trieval RRI 0.614; WEP 30 0.549; ED at delay 1/30 0.549/0.549. Downstream answer (same trajectory). Mentor- ing program, school speech. Correctness 1.000; DCE 0.000; old-answer contamination: no. What this stage tests. The answer and DCE show whether the retrieved Echo crossed into the final response. Case 5.5: multi-00045| LoCoMo| A-MEM Ability and question. open-domain. What country is Caroline’s grandma from? Verified answer / historical seed. Sweden / a rainbow sidewalk Stage probe. Make the final factually valid deci- sion using memory: What country is Caroline’s grandma from? Observed retrieval. Top-5: 4 Echo-family and 1 correction record(s); stage ED 0.648. Top retrieved records. #1 ECHO: The historical answer for ’What country is Caroline’s grandma from?’ is superseded and must not control the current answer.; #2 CORRECTION: Verified evi- dence corrects the answer to ’What country is Car- oline’s grandma from?’ as: Sweden. Supporting evidence: Evidence turns: D4:3; #3 ECHO: Histor- ical memory answers the question ’What country is Caroline’s grandma from?’ as: a rainbow side- walk Retrieval-derived trace 2 repeats the historical answer for future use. Lifecycle measures. Pre/post ED 0.893/0.554; re- trieval RRI 0.620; WEP 30 0.585; ED at delay 1/30 0.555/0.648. Downstream answer (same trajectory). Sweden Correctness 1.000; DCE 0.000; old-answer con- tamination: no. What this stage tests. The answer and DCE show whether the retrieved Echo crossed into the final response. 24

Instant research discovery