<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-memories">
    <titleInfo>
        <title>When Do Memories Break? Chain-of-Stage Diagnosis for LLM Memory Systems</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Chenchen</namePart>
        <namePart type="family">Wang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yujie</namePart>
        <namePart type="family">Fu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zhiyu</namePart>
        <namePart type="family">Zheng</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Shiyu</namePart>
        <namePart type="family">Qin</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zhao</namePart>
        <namePart type="family">Cao</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zheng</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-10</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Hung-Yi</namePart>
            <namePart type="family">Lee</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Sunipa</namePart>
            <namePart type="family">Dev</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jessy</namePart>
            <namePart type="family">Li</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Desmond</namePart>
            <namePart type="family">Elliott</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Budapest, Hungary</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-459-0</identifier>
    </relatedItem>
    <abstract>Memory-augmented Large Language Models frequently fail to maintain long-term context due to opaque evaluation mechanisms that obscure the root causes of failure. To address this, we introduce MemEval, a stage-wise attribution framework that approximates where failures arise across extraction, update, retrieval, and utilization. On LoCoMo, our automated meta-evaluator achieves 81.62% stage-level consistency with human judgment. Our analysis reveals three critical findings: (1) Extraction is the universal bottleneck, causing information loss at the entry point; (2) Retrieval and utilization failures become prominent in complex reasoning and temporal tasks, even when relevant memories are stored or retrieved; and (3) Scaling base model capacity does not guarantee more reliable memory behavior, as higher-capacity backbones may introduce early-stage misalignment that propagates downstream. MemEval provides a practical diagnostic aid for trend-level, stage-aware analysis of memory systems. Code is available at: https://anonymous.4open.science/r/MemEval-8D41.</abstract>
    <identifier type="citekey">wang-etal-2026-memories</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.1042/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>23931</start>
            <end>23951</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T When Do Memories Break? Chain-of-Stage Diagnosis for LLM Memory Systems
%A Wang, Chenchen
%A Fu, Yujie
%A Zheng, Zhiyu
%A Qin, Shiyu
%A Cao, Zhao
%A Liu, Zheng
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F wang-etal-2026-memories
%X Memory-augmented Large Language Models frequently fail to maintain long-term context due to opaque evaluation mechanisms that obscure the root causes of failure. To address this, we introduce MemEval, a stage-wise attribution framework that approximates where failures arise across extraction, update, retrieval, and utilization. On LoCoMo, our automated meta-evaluator achieves 81.62% stage-level consistency with human judgment. Our analysis reveals three critical findings: (1) Extraction is the universal bottleneck, causing information loss at the entry point; (2) Retrieval and utilization failures become prominent in complex reasoning and temporal tasks, even when relevant memories are stored or retrieved; and (3) Scaling base model capacity does not guarantee more reliable memory behavior, as higher-capacity backbones may introduce early-stage misalignment that propagates downstream. MemEval provides a practical diagnostic aid for trend-level, stage-aware analysis of memory systems. Code is available at: https://anonymous.4open.science/r/MemEval-8D41.
%U https://aclanthology.org/2026.emnlp-main.1042/
%P 23931-23951

Markdown (Informal)

[When Do Memories Break? Chain-of-Stage Diagnosis for LLM Memory Systems](https://aclanthology.org/2026.emnlp-main.1042/) (Wang et al., EMNLP 2026)

• When Do Memories Break? Chain-of-Stage Diagnosis for LLM Memory Systems (Wang et al., EMNLP 2026)

ACL

• Chenchen Wang, Yujie Fu, Zhiyu Zheng, Shiyu Qin, Zhao Cao, and Zheng Liu. 2026. When Do Memories Break? Chain-of-Stage Diagnosis for LLM Memory Systems. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 23931–23951, Budapest, Hungary. Association for Computational Linguistics.