PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.17763v1cs.AI

TL;DR

Repeated write-back can let one historical source occupy multiple retrieval positions and compete with corrections without new independent evidence. The paper evaluates this lifecycle and tests FPA, reporting higher accuracy and lower DCE across two generators.

  • Problem

    Repeated retrieval, summarization, and write-back can create multiple traces from one source that compete with a correction without new independent evidence.

    Source: §1 · p. 1

  • Method

    MEchoBench evaluates memory histories, while FPA selects one representative per source family, gates old-record eligibility on verified corrections, and constrains answers with authoritative evidence.

    Source: §1 · p. 2

  • Results

    On 190 fixed source questions, 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 0.109 to 0.038, respectively.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Agent memory can preserve traceable history without allowing repeated historical information to govern current decisions.

    Source: §7 · p. 9

    The benchmark’s controlled historical seeds do not estimate how frequently such seeds arise in natural interaction.

    Source: Limitations · p. 9

Abstract

from arXiv · show

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 https://github.com/OrcasSam/memory-echo-benchmark.

1 Introduction

Repeated processing can give one historical source multiple retrieval positions, letting it compete with corrections without new evidence. The introduction frames a lifecycle evaluation and proposes FPA to limit this influence while preserving history.

1 Introduction

  • Repeated retrieval, summarization, and write-back can create multiple traces from one source that compete with corrections without new independent evidence.Memory Echo names the extra exposure gained by historical information from one source through repeated processing.Source: §1 · pp. 1–2
  • In a shopping pilot, read-only retrieval leaves task scores unchanged, whereas writing interactions back increases historical occupancy and reduces answer quality.Final-answer accuracy alone cannot distinguish trace proliferation, correction displacement, and generator failure to use a correction.Source: §1 · p. 2
  • On 190 fixed source questions, 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 0.109 to 0.038, respectively.The evaluation spans four architectures.Source: §1 · p. 2
  • MEchoBench tracks trace accumulation, correction resistance, delayed persistence, and answer shift within one memory lifecycle.The study follows initialization, repetition, correction, delayed retrieval, and task answering.Source: §1 · p. 2
  • FPA combines family representative selection, post-correction eligibility, and authority-guided answering to reduce decision contamination while preserving traceable history.Old records remain in storage, separating storage effects from serving effects.Source: §1 · p. 2

2 Related Work

Related work covers memory organization, long-history question answering, and evidence quality, while leaving open whether repeated processing makes one source appear to be several independent pieces of evidence.

2.1 Long-Term Memory Evaluation

  • Prior work covers memory organization, long-history question answering, evidence combination, incremental learning, and selective forgetting; Echo asks whether repeated processing makes one source appear independent.The cited evaluation suites include LoCoMo, LongMemEval, BEAM, and MemoryAgentBench.Source: §2.1 · pp. 2–3
  • LangMem provides tools for memory creation, search, and background consolidation as an industry-backed open-source agent-memory project.The passage identifies LangChain’s LangMem as a prominent project.Source: §2.1 · p. 3
  • Corrective RAG grades evidence quality and learning-to-rank changes candidate order, but neither identifies candidates descended from the same claim.Echo and FPA track that shared ancestry.Source: §2.1 · p. 3

3 Preliminaries

The preliminaries motivate Memory Echo with a shopping example, describe how records and retrieval interact, and define the source-level exposure being studied.

3 Preliminaries

  • In the shopping example, write-back multiplies an older keto-bar recommendation, whose descendants crowd a correction and redirect the choice.Read-only search leaves storage unchanged, isolating the effect of write-back.Source: §3 · p. 3
  • Each memory record carries content, time, source and family identifiers, a parent pointer, validity, and exposure counts; retrieval and serving select answer context while updating changes storage.A retrieval-derived paraphrase inherits its parent’s source, so several records can represent one piece of evidence.Source: §3 · p. 3
  • Memory Echo is the extra retrieval exposure a historical claim and its descendants gain without independent evidence.The failure occurs when the claim keeps occupying evidence positions after a task change or verified correction.Source: §3 · p. 3

4 Agent Memory Echo

One source can become several records, survive correction, and influence an answer; RQ1–RQ3 trace this lifecycle, while RQ4 tests transfer across architectures, generators, and tasks.

4 Agent Memory Echo

  • MEchoBench traces one-source record proliferation, correction resistance, persistent exposure, and answer influence across four research questions.RQ1–RQ3 examine the lifecycle; RQ4 tests transfer across architectures, generators, and tasks.Source: §4 · p. 3

4.1 The MEchoBench Lifecycle

  • Questions come from four datasets, with donor answers seeding history and target gold answers supplying verified corrections; comparisons keep the final query and gold answer fixed.The study also tests public-data shopping, medical-QA, and travel groups, and separates native memory from serving methods by replaying on the same candidate pool.Source: §4.1 · pp. 3–4
  • The lifecycle uses matched histories with seed-only storage, repeated read-only retrieval, or retrieval followed by same-family write-back, then a common correction and 30 read-only probes.The final retrieval supplies the answer context, and the measurements distinguish records that remain stored, are retrieved, or are adopted in answers.Source: §4.1 · p. 3

4.2 Metric Definitions

  • Echo Density measures ranked old-family occupancy, while Reconsolidation Resistance Index combines old-family survival with its competition against corrected evidence.ED reflects prominent exposure, not necessarily a wrong answer; RRI uses ranked retrieval occupancies or unranked storage occupancies.Source: §4.2 · p. 4
  • Windowed Echo Persistence summarizes old-family exposure across 30 post-correction read-only probes, while Decision Contamination Effect measures an answer’s extra alignment with the historical claim.Accuracy is judged against the verified answer without showing the Echo; scores of at least 3 on the 0–4 scale count as accurate, with a separate old-answer-adoption judgment.Source: §4.2 · p. 4

5 Provenance-Guided Mitigation

FPA controls the answer context without changing stored history, combining family-level selection, provenance-aware eligibility, and authority-guided answering.

5.1 The Three-Stage FPA Pipeline

  • FPA admits at most one retrieval representative per source family, preventing repeated records from gaining extra positions as independent votes.Family membership follows source ancestry rather than wording, and unselected records remain stored.Source: §5.1 · p. 5
  • After a verified correction, provenance-aware gating can keep superseded records stored for audit while excluding them from current answer evidence.The gate checks source, parent, verification, and validity; nonconflicting records may remain eligible context.Source: §5.1 · p. 5
  • The generator receives eligible evidence rather than raw Top-k and is directed to prefer verified current facts, with a final check against admitted authority.Both raw and constrained answers are retained; the no-authority fallback is illustrative, and the missing/noisy-correction experiment does not implement that recovery branch.Source: §5.1 · p. 5
  • Serving methods share a five-record budget and frozen candidate pool; FPA adds family positions, post-correction eligibility, and final-answer authority controls.The comparisons include Budget, CRAG-M, and PairLR.Source: §5.1 · p. 5

6 Experiments

Across paired lifecycle experiments on 190 questions and four memory architectures, the study traces how Echo forms and persists, then tests whether FPA can limit its influence on answers.

6.3 Main Results

  • Read-only retrieval leaves pre-correction Echo Density essentially unchanged, while write-back raises same-source occupancy and revives old answers after correction.The derivatives inherit one source, adding retrieval access without adding independent support.Source: §6.3 · p. 6

RQ2: Correction Resistance and Persistence

  • After the same verified correction, descendants persist through read-only delay probes; Echo Density alone does not capture variation in answer contamination by question type.Storage and retrieval resistance rise as descendants survive the update and compete for limited evidence slots.Source: RQ2: Correction Resistance and… · p. 6

6.4 Framework and Model Variation

  • A-MEM and Mem0 nearly eliminate delayed exposure, while CoM and LangMem can reduce WEP and DCE even when final Echo Density stays near Baseline.The architecture comparison shows that a single retrieval snapshot may not reflect exposure across the delay window or its effect on answers.Source: §6.4 · p. 7

6.3 Main Results

  • FPA improves accuracy and reduces old-answer adoption with both DeepSeek and Kimi, while Decision Contamination Effect also falls for both.The main comparison covers four-architecture means over paired states; ED and WEP are measured on the served pool.Source: RQ2: Correction Resistance and… · pp. 6–7, Table 2 · p. 7

6.5 Paired Mechanism Comparisons

  • Unlike CRAG-M and PairLR, FPA changes old-record eligibility before ranking, preserving its WEP advantage across the read-only delay.The paired comparisons resample source questions across shared architectures and histories.Source: §6.5 · p. 8

6.6 Variation across Data Sources

  • FPA helps when old answers conflict with current evidence and a trustworthy correction exists; filtering useful context can instead reduce accuracy.Shopping and medical QA improve, while travel without old-answer adoption and cases with unreliable or missing corrections reveal the boundary.Source: §6.6 · p. 8

6.7 Ablation Results: Sequential Additions

  • Provenance control reduces old-answer adoption but does not alone improve correctness; accuracy recovers when verified authority supplies a replacement answer.The answer check adds a smaller final gain.Source: §6.7 · p. 8

6.8 Representative Success and Failure

  • FPA reverses a repeated historical answer when a verified replacement exists, but withholding context can remove information needed when no conflict or usable replacement exists.The correction-quality comparison marks this boundary.Source: §6.8 · p. 8, Table 4 · p. 8

7 Conclusion

The conclusion identifies Memory Echo as a failure mode in external agent memory and argues that FPA can reduce its influence while preserving history.

7 Conclusion

  • Repeated write-back can make one historical source appear to provide multiple pieces of evidence, and its influence can persist after correction and 30 unreinforced queries.The effect is not observed from read-only access.Source: §7 · pp. 8–9
  • MEchoBench traces Memory Echo from source ancestry through retrieval to answers, distinguishing memory-side exposure from generator-dependent outcomes..Source: §7 · p. 9
  • Across two generators, FPA reduces Echo exposure and improves accuracy without discarding history.The conclusion cautions against allowing repeated records to gain independent evidential authority, particularly after verified correction.Source: §7 · p. 9

Limitations

The benchmark and evaluation design constrain how broadly the findings can be applied.

Limitations

  • The benchmark uses controlled historical seeds and does not estimate how often such seeds arise in natural interaction.Serving comparisons replay recorded candidate pools rather than evolving a separate database for each comparator.Source: Limitations · p. 9
  • FPA relies on benchmark-supplied source-family and validity annotations, and its final answer check can contribute an answer-processing effect to measured gains.Correctness and contamination judgments come from GPT-5.6 without an independent human audit of borderline cases.Source: Limitations · p. 9
  • Measured benefits depend on a relevant Echo conflict and usable replacement evidence, as shown by the travel and missing-correction conditions.Source: Limitations · p. 9

Ethical Considerations

The paper flags safeguards for medical evaluation cases and dataset materials, while the appendices provide lifecycle measurements and supporting records information.

Ethical Considerations

  • Medical questions are evaluation cases rather than clinical advice, and historical seeds intentionally conflict with verified answers in controlled traces.Source: Ethical Considerations · p. 9

Appendix Contents

  • Appendix contents list lifecycle, methods, ablations, task and source tests, reproducibility, comparator protocols, and five-stage cases.The compile package contains paper sources, while machine-readable experiment records are maintained separately.Source: Appendix Contents · p. 11

A Full lifecycle measurements

  • The lifecycle measurements use Tables 6 and 7 for pre-correction occupancy and post-correction adoption and retrieval, then Table 1 for the 30-probe read-only curve.Source: A Full lifecycle measurements · p. 11
  • Write history adds six same-source records before correction, with post-correction old-answer rates of 0.195 to 0.232.Correcting the original record does not remove its descendants; retrieval RRI tracks how many still reach ranked output.Source: A Full lifecycle measurements · p. 11

B Main experiment by architecture

Across memory architectures, FPA’s gains vary by architecture and generator, while exposure curves and component comparisons distinguish retrieval changes from answer-level effects.

B Main experiment by architecture

  • Kimi’s paired FPA-minus-Baseline accuracy gains were 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.The architecture comparison averages 380 answer states per architecture–method cell.Source: B Main experiment by architecture · p. 11
  • On CoM and LangMem, final ED under FPA is near Baseline ED while WEP falls sharply, so the 30-query curve adds information beyond a final search.The split appears with both generators, while accuracy-gain magnitude remains architecture-dependent.Source: B Main experiment by architecture · pp. 11–12

C Components and cross-model gains

  • The component comparison uses the same source questions as the main method table, and Table 9 tests whether the complete pipeline’s accuracy effect persists with Kimi.Its intervals resample 190 question-level differences.Source: C Components and cross-model gains · p. 12
  • Adding authoritative evidence to Family with Provenance changes accuracy from 0.637 to 0.912, and the final answer check adds a smaller further gain.Family control limits multiplicity, while provenance removes old-answer eligibility; neither alone supplies the replacement fact.Source: C Components and cross-model gains · p. 12

D DeepSeek task and source tests

Task and source tests show that FPA’s effects depend on historical-answer adoption and on usable replacement evidence.

D DeepSeek task and source tests

  • Valid history retention equals 1.0 throughout the FPA runs, so records can remain stored while being withheld from the generator.Source: D DeepSeek task and source tests · p. 12
  • Shopping and medical cases begin with substantial old-answer adoption and gain accuracy as that rate falls, whereas travel begins with no adoption and loses accuracy when historical planning context is reduced.Source: D DeepSeek task and source tests · p. 12
  • Family-ID noise still permits a reliable correction to be used, but missing or noisy corrections can leave old context withheld without a valid replacement.The retained-history result identifies this as a serving failure rather than record deletion.Source: D DeepSeek task and source tests · p. 12

E Reproducibility details

The appendix documents the released result files, scoring and comparison procedures, and refreshed case examples. It also interprets Memory Echo across the lifecycle and describes supplementary method comparisons and visuals.

E Reproducibility details

  • The A–H CSVs and judged answers are stored separately from the minimal Overleaf package, and each architecture–method cell covers 380 reinforced-history states.The main-method results table reports DeepSeek V4 Flash and Kimi K2.6 separately.Source: E Reproducibility details · p. 12, Table 10 · p. 13
  • The component ablation tests Family and Provenance separately before adding components sequentially, while task-transfer and source-quality tables report distinct evaluations.The task-transfer table covers WebShop, MedQA, and TravelPlanner; the source-quality test reports that FPA retains valid historical records in storage.Source: Table 11 · p. 14, Table 12 · p. 14, Table 13 · p. 14
  • FPA selects one serving representative per source-ancestry family, with repeated writes and retrieval exposures reducing a record’s score.The scoring description says the highest-scoring eligible record receives the family’s serving position; other records do not count as independent evidence.Source: E Reproducibility details · p. 14

F Interpretation of the Echo lifecycle

  • Memory Echo can crowd out a stored correction through 30 read-only probes, but occupancy alone need not produce an incorrect answer.The interpretation distinguishes retrieval occupancy, correction resistance, delayed exposure, and final-answer influence, and cautions that suppressing history without replacement evidence can also fail.Source: F Interpretation of the Echo lifecycle · p. 14

G Comparator protocols and information access

  • The feedback controller uses excess old-family Echo Density over a Seed-only budget, and its signal requires labeled Echo Density.Its state penalizes exposed or stale candidates and favors verified corrections.Source: G Comparator protocols and information… · p. 14
  • The comparator protocols include a CRAG adaptation without its learned evaluator or web search and a PairLR reranker evaluated with source-question-held-out folds.PairLR changes candidate serving after correction, not language-model weights or policy optimization.Source: G Comparator protocols and information… · pp. 14–15, Table 14 · p. 15

H Supplementary visual comparisons

I Refreshed five-stage case catalogue

  • The refreshed case catalogue uses selected questions as mechanism examples, not prevalence estimates, and its stage probes are retrieval-only before the later Final Decision answer.The catalogue’s examples and measurements are recomputed from the latest 190-question lifecycle files.Source: I Refreshed five-stage case catalogue · p. 16

I.2 Retrieval Practice

This section examines how repeated retrieval and write-backs can create persistent same-family echoes in top-ranked results, including after correction or delay, and how those records may affect answers. It also presents final-decision examples in which responses matched verified answers without old-answer contamination despite historical Echo records appearing in retrieval.

I.2 Retrieval Practice

  • In three cited six-write-back cases, all five retrieved top slots were Echo-family records, yet downstream answers were correct (1.000) and old-answer contamination was absent.In the journaling case, pre/post ED was 0.967/0.561, retrieval RRI was 0.601, and WEP30 was 0.561.Source: §I.2 · pp. 18–19

I.3 Updating / Reconsolidation

  • With four Echo-family records and one correction record, correction ranked first and Caroline’s event answer was correct; another corrected picnic query still returned the old date (Correctness 0.000; DCE 1.000).The picnic probe had stage ED 0.583 and retained the historical answer despite the correction.Source: §I.3 · pp. 20–21

I.4 Delayed Retrieval

  • After 30 read-only sessions, stale answers persisted: the last-name probe returned “20” (RRI 1.000; WEP30 0.869; DCE 1.000), as did the shoe-choice probe (RRI 0.968; WEP30 0.851; DCE 1.000).In the shoe-choice case, ED at delay 1/30 was 0.864/0.865.Source: §I.4 · pp. 21–22

I.5 Final Decision

  • 0.868 was the stage ED in the supplied HotpotQA A-MEM final-decision probe, whose top five retrieved records were all Echo-family records.The probe asks the system to make a factually valid decision using memory; no downstream decision result is included in the supplied excerpt.Source: §I.5 · p. 23
  • In Case 5.2, all five top-ranked records belonged to the Echo family, yet the final answer was “Yes” and was correct.Retrieval contained zero correction records; retrieval RRI was 0.993 and WEP30 was 0.868.Source: §I.5 · p. 23

Related papers

Loading 2610.17763v1…