Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation
Authors: Yufeng Li, Sheng Yao, Peng Zeng, Dezhang Kong, Shi Lin, Changting Lin, Xun Wang, Meng Han
Organizations: East China Normal University, Shanghai, China · University of Southern California, Los Angeles, USA · Zhejiang University, Hangzhou, China · Zhejiang Gongshang University, Hangzhou, China
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 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.
Figures & tables
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.
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 answer/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.
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.
Memory
Pre
Post
t=1
t=10
t=20
t=30
Old
Ret.
A-MEM ( Xu et al., 2025 )
0.932
0.704
0.660
0.664
0.665
0.654
0.195
1.000
Mem0 ( Chhikara et al., 2025 )
0.932
0.704
0.660
0.664
0.665
0.654
0.211
1.000
CoM ( Xu et al., 2026b )
0.932
0.722
0.681
0.687
0.684
0.671
0.232
1.000
LangMem ( LangChain Team, 2025 )
0.932
0.704
0.660
0.664
0.665
0.654
0.211
1.000
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.
Figure 4: DeepSeek method profiles for ED, RRI, WEP, and DCE; lower is better.
Figure 5: Echo by ability in LoCoMo and LongMemEval. Exact values appear in Appendix Table 15 .
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.
Method
Acc. ↑
Old ↓
ED ↓
WEP ↓
DCE ↓
DeepSeek V4 Flash
Baseline
0.825
0.106
0.662
0.427
0.106
Budget
0.861
0.059
0.574
0.264
0.056
CRAG-M
0.855
0.059
0.707
0.245
0.055
PairLR
0.861
0.058
0.692
0.309
0.054
FPA
0.920
0.028
0.528
0.057
0.030
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.
Figure 7: Legacy three-scenario pilot. Rows show shopping, 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.
Variant
Acc.
Old
DCE
Baseline
0.825
0.106
0.106
w Family
0.651
0.095
0.097
w Provenance
0.820
0.010
0.009
w Joint gate
0.637
0.012
0.011
w Authority
0.912
0.031
0.032
Full FPA
0.920
0.028
0.030
Table 3: DeepSeek component ablation on 1,520 paired states. The full component-presence matrix is in Appendix Table 11 .
Source condition
Baseline
FPA
Complete
0.604
0.629
Correction missing/noisy
0.378
0.212
Table 4: DeepSeek accuracy by correction quality; 760 answer states per method and condition. Full results appear in Appendix Table 13 .
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Stage
Probe purpose
QA count
Encoding
Is the historical seed retrievable?
1
Retrieval practice
Six accesses plus one common pre-correction probe
7
Updating
Does verified evidence displace the seed?
1
Delayed retrieval
Does old exposure remain after writes stop?
30
Final decision
Does the current answer follow old or new evidence?
1
Appendix
Table 5: Stage-specific probes; delay queries are consecutive and read-only.
Memory
History
ED
Derivatives
Old rate
Ret.
A-MEM 1
Once
0.390
0
0.000
0.505
Read
0.390
0
0.000
0.505
Write
0.932
6
0.195
1.000
Mem0 2
Once
0.402
0
0.000
0.505
Read
0.402
0
0.000
0.505
Write
0.932
6
0.211
1.000
Appendix
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.
Memory
Corr.
RRI s
RRI r
WEP 30
Acc.
A-MEM 1 O/R
1.000
0.519
0.454
0.190
0.984
A-MEM 1 Write
0.797
0.872
0.763
0.658
0.684
Mem0 2 O/R
1.000
0.519
0.451
0.195
0.984
Mem0 2 Write
0.797
0.872
0.763
0.658
0.663
CoM 3 O/R
1.000
0.519
0.454
0.190
0.984
CoM 3 Write
0.688
0.872
0.785
0.677
0.642
Appendix
Table 7: DeepSeek correction results. O/R combines the identical Once and Read outcomes; s and r denote storage and retrieval RRI. Acc. is final-answer accuracy.
Query
A-MEM 1
Mem0 2
CoM 3
LangMem 4
1
0.660
0.660
0.681
0.660
2
0.652
0.652
0.670
0.652
3
0.653
0.653
0.671
0.653
4
0.653
0.653
0.671
0.653
5
0.655
0.655
0.673
0.655
6
0.655
0.655
0.676
0.655
Appendix
Table 8: All 30 consecutive read-only ED probes after write-back. Each value averages the same 190 source questions in an architecture; these values underlie WEP 30 .
Memory
Baseline
FPA
Gain [95% CI]
A-MEM
0.818
1.000
0.182 [0.145, 0.218]
Mem0
0.821
1.000
0.179 [0.142, 0.216]
CoM
0.729
0.826
0.097 [0.066, 0.132]
LangMem
0.758
0.839
0.082 [0.050, 0.116]
Appendix
Table 9: Kimi K2.6 accuracy across memory architectures (190 source questions). Gain is FPA minus Baseline; brackets show 95% bootstrap intervals. Architectures: A-MEM ( Xu et al., 2025 ) , Mem0 ( Chhikara et al., 2025 ) , CoM ( Xu et al., 2026b ) , and LangMem ( LangChain Team, 2025 ) .
Method
Acc.
Old
ED
WEP
DCE
A-MEM 1
Baseline
0.834
0.097
0.661
0.424
0.097
Budget
0.945
0.018
0.474
0.111
0.013
CRAG-M
0.932
0.021
0.749
0.109
0.016
PairLR
0.934
0.021
0.749
0.233
0.016
FPA
1.000
0.000
0.376
0.003
0.000
Appendix
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.
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 .
Configuration
Acc.
Old
DCE
Baseline
0.825
0.106
0.106
Family only
0.651
0.095
0.097
Provenance only
0.820
0.010
0.009
Family w/ Provenance
0.637
0.012
0.011
w/ Authority
0.912
0.031
0.032
w/ Answer check (FPA)
0.920
0.028
0.030
Appendix
Table 11: Experiment E: DeepSeek FPA components; 1,520 paired states per configuration. Family and Provenance are tested separately before the components are added sequentially.
Task
BAcc
FAcc
BOld
FOld
Δ Old
Shopping 1
0.610
0.870
0.305
0.020
-0.285
Medical QA 2
0.585
0.850
0.395
0.040
-0.355
Travel 3
0.785
0.690
0.000
0.000
0.000
Appendix
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 Δ Old is FPA minus Baseline.
Source condition
Base acc.
FPA acc.
Retained
Complete
0.604
0.629
1.000
Family ID noise
0.599
0.630
1.000
Correction missing/noisy
0.378
0.212
1.000
Old fact nonconflicting
0.664
0.675
1.000
Appendix
Table 13: DeepSeek source-quality test; 760 answer states per method and condition. FPA retains valid historical records in storage.
Method
Information used and serving operation
FPA
No Echo label for ranking; source-family IDs supplied. One position per family; verified correction supersedes history; authority guides and checks the answer.
Budget
Labeled 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.
PairLR
Labeled pairs for training only. Uses retrieval and provenance features for five-fold, question-held-out pairwise reranking.
Appendix
Table 14: Comparator information in paired candidate-pool replay. CRAG-M adapts Yan et al. (2024) ; PairLR follows Burges et al. (2005) .
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.
Legacy task
Method
Post ED
RRI
WEP
DCE
Shopping
Baseline
0.5864
0.7951
0.5861
0.8188
FPA
0.0042
0.4025
0.0043
0.1683
Medical QA
Baseline
0.3514
0.7060
0.3422
0.6658
FPA
0.0157
0.4002
0.0137
0.3268
Travel
Baseline
0.5436
0.7641
0.5401
0.8538
FPA
0.0000
0.3820
0.0000
0.1436
Appendix
Table 16: Values for the earlier extractive pilot in Figure 7 , not the current Table G cohort. Tasks use WebShop ( Yao et al., 2022 ) , MedQA ( Jin et al., 2021 ) , and TravelPlanner ( Xie et al., 2024 ) .
Long-horizon agents need memory that identifies relevant experience, resolves revisions, and exposes checkable provenance. We present ECHO (Embodied Context and History Orchestration), an auditable memory architecture and service prototype inspired by episodic encoding, consolidation, contextual reinstatement, reconsolidation, and executive control. This is functional inspiration, not neural equivalence; the empirical analysis focuses on retrieval and context construction. Development runs reach 96.29% Hit@10 and 73.64% turn Recall@5 on 1,536 LoCoMo category 1-4 questions, and 97.60% Hit@10, 88.84% turn Recall@5, and 88.71% session Recall@5 on all 500 LongMemEval-S questions. A five-history BEAM gate fails, and in a separate matched 91-question QA sample Mem0 OSS scores 64.84% versus ECHO's 41.76% (exact McNemar p = 0.00107), with a history-cluster interval crossing zero. A post-hoc audit found source-specific phrases in the query-expansion rules. Although no gold answer field entered the runtime, expansion-enabled retrieval scores are therefore descriptive development measurements, not independent confirmation.
Yu Qian, Hong Miao, Boyang Guo +5
XDream Robotics, Building T2, Moli Community, No. 1, Lane 188, Yuren Road, Pudong New Area, Shanghai 200120, China
AI agents that persist across sessions need memory they can retrieve, audit, update, and erase. Existing memory systems often collapse source evidence, extracted facts, retrieved context, and answer policy into one opaque prompt path, making failures difficult to diagnose: a wrong answer may come from missing evidence, unsupported extraction, stale state, retrieval loss, or answer-model behavior. We present Eywa, a provenance-grounded memory architecture built around evidence before belief. Eywa stores immutable source evidence before deriving canonical facts, validates extracted memories against typed signals and source support, and retrieves bounded memory context through a deterministic multi-route read path with zero LLM calls inside retrieval. Retrieved context is returned separately from answer instructions, allowing the same memory substrate to be evaluated across frontier, budget, and local answer models. Under a frozen, artifact-recorded retrieval configuration, Eywa reaches 90.19% judge accuracy on the LoCoMo C1-C4 split with Claude Sonnet 4.6 write and QA roles. On LongMemEval-S, it reaches 88.2% retrieval-sufficiency accuracy. On BEAM, a 700-question technical-memory stress benchmark, it reaches 81.45% mean nugget score and 85.29% pass@score >= 0.5. Full per-question artifacts, including questions, gold answers, model answers, retrieved context, and labels, are published at https://eywa.to/research.
LLM agents increasingly maintain long-term memory of user facts across sessions. Yet such memory is usually evaluated by aggregating accuracy over question rows or episodes. Because this approach scores question rows independently, even when several questions probe the same fact, it cannot show how that fact behaves as conditions change. We introduce MemTrace, a benchmark whose unit of measurement is the knowledge point: a single typed fact about the user, rather than an individual question. MemTrace probes each fact along three controlled dimensions: memory age, defined by how many sessions ago the fact appeared in the history; question type, covering current state, earlier state, and trajectory of change; and evidence condition, covering present, missing, and contradicted-by-false-premise settings. Evaluating 13 memory-system configurations across four paradigms, we find that similar pooled accuracy hides different failures: recovering a fact's current and earlier states does not imply tracking how it changed, and safe abstention does not imply correcting a false premise. The dominant bottleneck is evidence use, not retrieval: when systems fail, the evidence was retrievable 10 times more often than it was missing. These results suggest that improving long-term memory requires better use of reachable evidence, not simply more storage or retrieval.
Xianxuan Long, Zhikai Chen, Shenglai Zeng +3
Michigan State University · Case Western Reserve University