ES-MemEval: Benchmarking Conversational Agents on Personalized Long-Term Emotional Support

Personas, identity, and agents2026arXivApproved editorial review

Authors: Tiantian Chen, Jiaqi Lu, Ying Shen, Lin Zhang

Keywords: Long-term conversational memory, Emotional support agents, Benchmark evaluation, Retrieval-augmented generation, LLM-as-judge, Synthetic data provenance

Source: Open primary source (opens in a new tab)

4
Authors
16
Findings
21
Limitations
11
Evidence

Editorial summary

English

ES-MemEval proposes evaluating longitudinal conversational memory in emotional-support agents through five capabilities: information extraction, temporal reasoning, conflict detection, abstention, and user modeling. It does not study psychological personality, synthetic identity, or therapeutic efficacy; its defensible construct is retrieval and integration of personal information dispersed across conversations. The associated EvoEmo resource contains 18 virtual profiles, 401 sessions, 9,368 turns, 446 events, 2,300 observations, 125 summarization tasks, and 34 dialogue scenarios, with a mean history span of 447.72 days. Profiles are assembled from multiple ESConv seed dialogues; GPT-4o expands the timelines twice and generates later sessions that human annotators refine. The paper reports 1,209 QA questions, but the official JSON contains 1,427: 309 information-extraction, 284 temporal, 267 conflict, 261 abstention, and 306 user-modeling items. The released loader iterates all 1,427 and no filter, split, or evaluated-ID list identifies the 1,209-item paper sample, so QA tables cannot be tied to the downloadable benchmark. In the reported results, Mistral-24B moves with RAG from 15.5 to 18.8 F1, 47.4 to 50.4 BERTScore, and 1.01 to 1.27 under the GPT-4o judge. RAG does not improve everything: GPT-4o overall F1 falls from 26.6 to 23.9, BERTScore stays at 54.2, and abstention judge score falls from 1.67 to 1.30 even as its overall judge score rises from 1.25 to 1.33. RAG therefore helps conditionally by model, task, metric, and k, not uniformly. Every published summarization RAG row improves, with GPT-4o+RAG at ROUGE-L 22.2, event F1 49.4, and judge score 2.93. In dialogue generation, full history and RAG improve observation recall over no-memory, but GPT-4o Full and RAG receive 5/5/5 from GPT-4o for memory, personalization, and emotional support, creating a same-family circularity and ceiling. The three dimensions are also defined by one rubric around past-experience references and correlate almost perfectly in Table 8: pairwise rho 0.984-0.987. This is agreement with one judge/prompt, not independent evidence of emotional or clinical benefit. The artifact audit found defects that directly affect metrics. QA and summaries clean outputs with strip("Answer:"), which removes any matching edge characters rather than a prefix: `Answer: Peter` becomes `Pet`, and `Answer: Unknown` becomes `Unkno`. Because 258 of 261 abstention gold answers are `Unknown`, a correct unpunctuated answer can be scored as wrong. F1 uses sets rather than token multisets and crashes on empty strings. NDCG discounts the first result as 0.5, candidate sampling can call random.sample with a negative k, and the vector store does not retain IDs from its first load. Four summarization executables, Mistral-8B base/RAG, Phi-3 base, and Mistral-24B RAG, actually instantiate QaExperiment, so they do not generate four Table 6 rows. No outputs, generations, judge logs, or aggregatable CSVs are released, and the result organizer points to private `/home/lab509/...` paths; the tables cannot be recomputed. The JSON also needs repair: 37 QA evidence references are invalid, 3 questions duplicate evidence, 2 non-abstention questions have no evidence, and 7 question-text pairs are duplicated. Provenance contradicts a categorical ethics statement. Eighty-five sessions retain ESConv IDs, 58 reproduce complete dialogues after text normalization, and 2,252 of 9,368 utterances match the source corpus; `esc1198` appears under two profiles. ESConv is emotional-support role-play written by crowdworkers, not clinical records or natural therapy conversations, but it is human-authored conversational text. EvoEmo is therefore synthetic/augmented, not exclusively generated. Zenodo also labels the entire archive CC-BY-4.0 although it incorporates ESConv material distributed as CC-BY-NC-4.0 and described for academic research; provenance and licensing require clarification, without this audit offering a legal conclusion. Human validation is useful but narrow: only Mistral-24B Full/RAG, with 50 QA, 40 summaries, and 30 dialogues. For DG RAG, kappa 0.19 and rho 0.20 coexist with 86.7% exact agreement because of ceiling concentration, which does not validate fine ranking or GPT-4o self-judging. The faithful conclusion is that ES-MemEval offers an interesting longitudinal benchmark architecture and preliminary evidence that explicit memory and RAG can help in several conditions. Its numbers should remain provisional until the exact 1,209-question split is released, evaluation code is repaired, raw outputs are published, evidence is validated, ESConv lineage/licenses are documented, and broader independent human evaluation is performed. It does not establish synthetic personality, safe therapy, or production readiness.

Español

ES-MemEval propone evaluar memoria conversacional longitudinal en agentes de apoyo emocional mediante cinco capacidades: extracción de información, razonamiento temporal, detección de conflictos, abstención y modelado del usuario. No estudia personalidad psicológica, identidad sintética ni eficacia terapéutica; su constructo defendible es la recuperación e integración de información personal dispersa a lo largo de conversaciones. El recurso asociado, EvoEmo, contiene 18 perfiles virtuales, 401 sesiones, 9.368 turnos, 446 eventos, 2.300 observaciones, 125 tareas de resumen y 34 escenarios de diálogo, con una historia media de 447,72 días. Los perfiles se construyen a partir de varios diálogos semilla de ESConv; GPT-4o expande dos veces las líneas temporales y genera sesiones posteriores que anotadores humanos refinan. El paper declara 1.209 preguntas QA, pero el JSON oficial contiene 1.427: 309 de extracción, 284 temporales, 267 de conflicto, 261 de abstención y 306 de modelado. El loader publicado recorre las 1.427 y no existe filtro, split ni lista de IDs que identifique las 1.209 evaluadas, por lo que las tablas QA no se pueden vincular al benchmark descargable. En los resultados declarados, Mistral-24B pasa con RAG de 15,5 a 18,8 F1, de 47,4 a 50,4 BERTScore y de 1,01 a 1,27 en el juez GPT-4o. Sin embargo, RAG no mejora todo: GPT-4o baja de 26,6 a 23,9 F1 global, mantiene 54,2 BERTScore y su puntuación de abstención cae de 1,67 a 1,30 aunque el juez global sube de 1,25 a 1,33. En resumen, RAG ayuda de forma dependiente del modelo, tarea, métrica y k; no de manera uniforme. En summarization todas las filas RAG publicadas mejoran, con GPT-4o+RAG en ROUGE-L 22,2, event F1 49,4 y juez 2,93. En generación, historia completa y RAG aumentan recuerdo de observaciones frente a no-memory, pero GPT-4o Full y RAG obtienen 5/5/5 de GPT-4o en memoria, personalización y apoyo, una circularidad de familia con techo. Las tres dimensiones están además definidas por el mismo rubric en torno a referencias al pasado y correlacionan casi perfectamente en Table 8: rho 0,984-0,987 entre pares. Esto mide concordancia con un juez y prompt concretos, no beneficio emocional o clínico independiente. La auditoría del artefacto encontró defectos que afectan directamente a los números. QA y summaries limpian la salida con strip("Answer:"), que elimina caracteres de ambos extremos, no un prefijo: `Answer: Peter` se vuelve `Pet` y `Answer: Unknown` se vuelve `Unkno`. Como 258 de 261 respuestas de abstención tienen gold `Unknown`, una salida correcta sin puntuación puede contabilizarse como error. El F1 usa conjuntos en vez de multiconjuntos y falla con strings vacíos. El NDCG descuenta la primera posición como 0,5, el muestreo puede llamar random.sample con k negativo y el vector store no registra IDs de su primera carga. Cuatro ejecutables de resumen, Mistral-8B base/RAG, Phi-3 base y Mistral-24B RAG, ejecutan en realidad QaExperiment, de modo que no generan cuatro filas de Table 6. No se liberan outputs, generaciones, logs de jueces ni CSV agregables, y el organizador apunta a rutas privadas `/home/lab509/...`; las tablas no pueden recalcularse. La calidad del JSON también exige reparación: 37 referencias de evidencia QA son inválidas, 3 preguntas repiten evidencias, 2 preguntas no-abstention no tienen evidencia y hay 7 pares de texto duplicado. La procedencia contradice una afirmación ética categórica. Ochenta y cinco sesiones conservan IDs ESConv, 58 reproducen diálogos completos tras normalización textual y 2.252 de 9.368 utterances coinciden con el corpus fuente; `esc1198` aparece en dos perfiles. ESConv es role-play de apoyo emocional escrito por crowdworkers, no expedientes clínicos ni conversaciones terapéuticas naturales, pero sí texto conversacional humano. Por eso EvoEmo es sintético/aumentado, no exclusivamente generado. Además, Zenodo marca todo el archivo CC-BY-4.0 aunque incorpora material ESConv CC-BY-NC-4.0 y descrito para investigación académica; falta aclarar procedencia y licencia, sin que esta auditoría emita una conclusión legal. La validación humana es útil pero estrecha: solo Mistral-24B Full/RAG, con 50 QA, 40 summaries y 30 diálogos. En DG RAG, kappa 0,19 y rho 0,20 conviven con 86,7 % exacto por efecto techo, lo que no valida ranking fino ni self-judging de GPT-4o. La conclusión fiel es que ES-MemEval ofrece una arquitectura de benchmark longitudinal interesante y evidencia preliminar de que memoria explícita y RAG pueden ayudar en varias condiciones. Sus cifras deben considerarse provisionales hasta publicar el split exacto de 1.209 preguntas, corregir el evaluador, liberar outputs, validar evidencias, documentar ESConv/licencias y ampliar evaluación humana independiente. No demuestra personalidad sintética, terapia segura ni preparación para producción.

Research question

To what extent can LLMs with full history or retrieved memory extract, order, reconcile, abstain, and model implicit and changing personal information across multi-month emotional support conversations?

Method

The authors compose 18 profiles from several ESConv seed dialogues, expand the timelines twice with GPT-4o, and generate/refine longitudinal sessions. They build QA, summaries, and future dialogue scenarios. They compare Ministral-8B, Phi-3-Medium, Mistral-24B, GPT-3.5-turbo, and GPT-4o with full history and RAG BGE-M3/FAISS. QA uses F1, BERTScore, and a GPT-4o judge; summaries use ROUGE and extraction/GPT-4o judge; dialogue uses a GPT-4o-simulated seeker, a supporter model, observation judgments by Mistral-24B, and global ratings by GPT-4o. The audit read and visually reviewed the 12 pages, verified WWW '26/DOI, cloned the official commit and Zenodo, recounted all JSON, validated references, compared against official ESConv, audited license/provenance, formulas, scripts, configuration, outputs, tests, and reproducibility, and separated the paper's claims from what the artifact supports.

Sample: The paper describes 18 virtual users, 401 sessions, 23.4 turns, and 13,291.6 average tokens per conversation, 22.3 sessions per user, and 14.9 average months. The artifact confirms 18 users, 401 sessions, 9,368 turns, 2,300 observations, 446 events, 125 summaries, and 34 topics. It contains 1,427 QA, not the 1,209 of Table 2: 309 IE, 284 TR, 267 CD, 261 abstention, and 306 UM. The profiles are 13 Americans, 2 British, 2 Singaporeans, and 1 Australian; 11 women and 7 men; ages 17-40. Eighty-five sessions point to ESConv and 58 are complete normalized matches. Human reliability uses only Mistral-24B Full/RAG outputs and samples of 50 QA, 40 summaries, and 30 dialogues.

Findings

  • Mistral-24B+RAG improves over base on QA: 15.5 to 18.8 F1, 47.4 to 50.4 BERTScore, and 1.01 to 1.27 judge.
  • GPT-4o+RAG drops 26.6 to 23.9 F1, maintains 54.2 BERTScore, and improves 1.25 to 1.33 judge; its abstention judge drops 1.67 to 1.30.
  • Session top-8 achieves Recall 81.7 and published NDCG 62.4, but performs worse downstream than top-2/top-4 on main metrics.
  • On summaries, GPT-4o+RAG reaches ROUGE-L 22.2, event F1 49.4, and judge 2.93.
  • Full/RAG improve observation metrics over no-memory, but GPT-4o self-rates 5/5/5 on Full and RAG.
  • The three dimensions of Table 8 have rho 0.975-0.987 between pairs under a common rubric linked to the past.
  • The JSON has 1,427 QA versus 1,209 declared, and the public code processes the 1,427 without a published filter.
  • There are 37 invalid QA evidences, 3 lists with duplicates, 2 non-abstention questions without evidence, and 7 repeated question pairs.
  • 258/261 abstention gold are Unknown, and the cleaner converts Answer: Unknown without punctuation into Unkno.
  • Four summarization scripts run QaExperiment and cannot produce their Table 6 rows.
  • The published NDCG applies an incorrect discount, the candidate sampler can receive negative k, and all_documents omits the first vector load.
  • F1 uses sets and can raise ZeroDivisionError with empty strings.
  • No outputs/results/logs are published; the aggregator uses private paths, so the tables are not re-auditable.
  • 85 sessions come from ESConv, 58 match completely, and 2,252 utterances match human source text.
  • Zenodo marks the file CC-BY-4.0 despite incorporating ESConv CC-BY-NC-4.0/academic research; the license requires clarification.
  • The 119 Python scripts parse and the JSON is valid; there is no CI, automated suite, released results, or LICENSE on GitHub.

Limitations

  • The construct is conversational memory, not psychological personality or synthetic identity.
  • There are only 18 profiles, culturally narrow, in English, and predominantly American.
  • One profile is 17 years old, and the domain contains sensitive emotional health content.
  • Most of the longitudinal corpus is synthetic/augmented and may reflect GPT-4o's style, biases, and artificial consistency.
  • The claim of not containing real user conversations omits 85 ESConv sessions of human role-play.
  • These are not clinical data or natural therapeutic conversations, so they do not validate clinical performance either.
  • The evaluated benchmark of 1,209 questions is not identified within the 1,427 artifact.
  • Cleaning, F1, NDCG, sampling, vector store, and summary script defects contaminate reproducibility and metrics.
  • The 37 invalid evidence IDs prevent perfect tracing of supervision.
  • RAG does not improve uniformly and can harm abstention or performance with larger k.
  • GPT-4o judges GPT-4o on QA, summaries, and dialogues; family and style bias may exist.
  • Global ratings mix memory, personalization, and support through dependent definitions.
  • The 5/5 ratings and low dispersion produce ceiling effects and reduce correlation/ranking validity.
  • Human validation only covers two Mistral-24B conditions and small samples.
  • DG RAG has kappa 0.19 and rho 0.20; high exact agreement does not recover fine discrimination.
  • There are no confidence intervals, seed variation, or statistical tests for most comparisons.
  • No outputs, exclusions, per-sample failures, or judge logs are released for reanalysis.
  • The archive license does not clearly separate original, generated, and ESConv-derived content.
  • There is no repository LICENSE, CI, tests, complete lockfile, or GPU/API-free smoke test.
  • The paper declares use of A100 80GB; reproducing everything requires large local models and a commercial API.
  • The resource is marked for research only and not for counseling or clinical use.

What the study does not establish

  • It does not demonstrate synthetic personality, persistent identity, or psychological traits of the model.
  • It does not demonstrate that the evaluated agents are safe or effective therapists.
  • It does not demonstrate improvement in mental health, well-being, or real quality of support.
  • It does not publicly identify the 1,209 questions used in results.
  • It does not demonstrate that the tables can be reproduced with the published commit.
  • It does not demonstrate that RAG helps all models, metrics, capabilities, or values of k.
  • It does not demonstrate that the GPT-4o judge is independent or impartial when evaluating GPT-4o.
  • It does not demonstrate that 5/5 represents perfect quality outside the rubric.
  • It does not demonstrate that observation recall penalizes all invented memories.
  • It does not demonstrate that abstention fails without the answer cleaner bias.
  • It does not demonstrate that EvoEmo lacks human dialogue: it incorporates ESConv crowdworker role-play.
  • It does not demonstrate that the entire archive is reusable without restrictions under CC-BY-4.0.
  • It does not demonstrate cultural, multilingual, clinical, or production generalization.

Traceability

Scope: Full text

Version: arXiv:2602.01885v1, submitted 2 February 2026, 12 pages; WWW '26 DOI 10.1145/3774904.3792143; official repository/tag commit 692624208acc077b8867698c1d6fcd998dee641a; Zenodo 10.5281/zenodo.18338564

Consulted source: https://arxiv.org/abs/2602.01885v1

Review: Codex complete bilingual fidelity pass using the full 12-page arXiv v1 manuscript, all-page visual inspection, WWW '26/DOI metadata, official GitHub and Zenodo artifact audit, exact released-dataset recount, evidence-reference validation, official ESConv lineage comparison, license/provenance review, code/metric/reproducibility audit and independent Table 8 correlation checks; summaries written from full evidence rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Ministral-8B-Instruct-2410
  • Phi-3-Medium-128k-Instruct
  • Mistral-Small-3.1-24B-Instruct-2503
  • GPT-3.5-turbo
  • GPT-4o
  • BGE-M3 embeddings

Instruments and metrics

  • ES-MemEval five-capability benchmark
  • Full-history and no-memory conditions
  • BGE-M3 plus FAISS retrieval
  • Token F1 and BERTScore
  • ROUGE-1, ROUGE-2 and ROUGE-L
  • GPT-4o LLM-as-judge
  • Mistral-24B observation relevance and usage judgments
  • Recall@k and published NDCG@k
  • Weighted Cohen kappa, Spearman correlation and MAD

Data used

  • EvoEmo released JSON
  • ES-MemEval QA benchmark
  • ES-MemEval summarization benchmark
  • ES-MemEval dialogue-generation scenarios
  • ESConv crowdworker role-play source corpus

Evidence and location

  • Construct, creation of EvoEmo, and tasks: arXiv v1 pages 1-4, Sections 1-4, Figures 1-2 and Table 2
  • Models, RAG, and QA/retrieval/context results: arXiv v1 pages 4-6, Sections 5-6.1 and Tables 3-5
  • Summarization, dialogue, and global ratings: arXiv v1 pages 6-8, Sections 6.2-7 and Tables 6-8
  • Ethics, limitations, annotators, and human reliability: arXiv v1 pages 8-12, Appendices A-D and Table 9
  • Actual counts, evidence integrity, and demographics: Official data/evo_emo.json at commit 692624208acc077b8867698c1d6fcd998dee641a; independent full-file audit on 15 July 2026
  • ESConv provenance and session/utterance matches: Official EvoEmo JSON compared against thu-coai/ESConv official dataset snapshot; 85 source IDs, 58 complete normalized matches and 2,252 matching utterances
  • Artifact licenses and metadata: Zenodo 10.5281/zenodo.18338564 v1.0.0 metadata, official GitHub tag v1.0.0 and thu-coai/ESConv dataset card/license
  • Answer cleaning, F1, retrieval, and summary script defects: src/lib/qa/qa_experiment.py, src/lib/sum/sum_experiment.py, vector_document_store.py, qa_retrieval_experiment.py and src/exe/sum at audited commit
  • Judge circularity, prompts, and dimension correlation: All executable Config classes, dg_experiment.py, Table 8 values and independent rank-correlation calculation
  • Comprehensive technical, scientific, and data audit: reports/verification/article-202-es-memeval-data-code-and-validity-audit.json
  • Complete visual inspection: All 12 pages of arXiv:2602.01885v1 rendered and visually inspected on 15 July 2026