paper-anon-2026/CoMem-Benchmark-2026

Dataset Card for CoMem

CoMem is a benchmark derived from PersonaMem and LongDialQA (released with the DialSim framework), specifically curated via a two-stage LLM-verification pipeline to isolate preference evolution and entity tracking in both dyadic and multi-party settings.

For comprehensive methodological details, evaluation results, and additional manual annotation analysis, please refer to the accompanying paper.

---

Dataset Statistics

Subset	# Contexts	# Questions	Avg Context Length
PersonaMem 32k	37	200	~26k tokens
PersonaMem 128k	110	322	~118k tokens
LongDialQA	15	723	~120k tokens

Each retained question is annotated at four temporal checkpoints (τ ∈ {0.25, 0.5, 0.75, 1.0}) plus a τ = 0 reference point, yielding checkpoint-conditioned ground-truth labels per question.

---

Synthetic Data Generation and Provenance

This dataset is a real-data mixture augmented through generative processes. It contains synthetically generated Chain-of-Thought (CoT) reasoning and corrected answers derived from original text sources.

1. Seed Data: The foundational questions, options, and original answers are sourced from PersonaMem and LongDialQA. The PersonaMem shared dialogue contexts (shared_contexts_32k.jsonl and shared_contexts_128k.jsonl) are used directly without modification, preserving the original PersonaMem context structure.

2. Generators: The generation process utilizes two distinct Large Language Models to mitigate model-specific bias: GPT-5 (gpt-5-2025-08-07) for the first round of correction and Claude-Sonnet-4.5 (claude-sonnet-4-5-20250929) for the second round.

3. Generation Process (Multi-stage Correction): A two-round pipeline is applied to verify and augment the data at four context-truncation checkpoints (τ = 0.25, 0.5, 0.75, 1.0). In Round 1, GPT-5 takes the truncated context, question, options, and original answer as input, verifies the original answer against the context, and generates a step-by-step reasoning path (CoT) to justify its conclusion. In Round 2, Claude-Sonnet-4.5 receives the same input package along with the Round 1 reasoning, re-verifies the answer, and outputs an independent reasoning path for each checkpoint.

4. Validation (Consistency Check): The synthetic outputs are strictly validated using a cross-round consistency check. A question is retained if and only if the answers provided by both LLM rounds perfectly agree across all checkpoints. This validation explicitly accommodates dynamic reasoning: if the correct answer naturally shifts as more context is revealed across checkpoints (e.g., the answer sequence is A, B, C, C at τ = 0.25, 0.5, 0.75, 1.0), the question is retained as long as both GPT-5 and Claude-Sonnet-4.5 reach the exact same consensus sequence. Inconsistent instances are discarded. Detailed manual annotation process is reported in the accompanying paper.

---

Dataset Structure & Usage

1. PersonaMem (32K & 128K)

Evaluated System	Questions File	Context File
All Systems (32K)	personamem_32k/questions_32k.csv	personamem_32k/contexts_32k.jsonl
All Systems (128K)	personamem_128k/questions_128k.csv	personamem_128k/contexts_128k.jsonl

The context files for PersonaMem (contexts_32k.jsonl and contexts_128k.jsonl) are unmodified copies of the original PersonaMem shared_contexts_*.jsonl files, redistributed under the upstream MIT license with the original copyright notice preserved (see "Source Datasets & Licensing" below).

2. LongDialQA (released with DialSim)

Evaluated System	Questions File (Shared)	Context File (System-Specific)
Full Memory	dialsim/questions_dialsim.csv	dialsim/context_full.jsonl
RAG Memory	dialsim/questions_dialsim.csv	dialsim/context_rag.jsonl
Mem0 & Memp	dialsim/questions_dialsim.csv	dialsim/context_alternate.jsonl

---

Data Use Cases

What this dataset measures: The dataset evaluates long-context LLM-based agents on two memory-related abilities: (1) tracking continual user preference evolution across multi-session dyadic dialogue, and (2) maintaining long-horizon entity tracking across multi-party dialogue. It operationalizes these via a checkpoint-conditioned QA protocol over 26K-128K-token contexts at five temporal cutoffs (τ ∈ {0, 0.25, 0.5, 0.75, 1.0}), where τ = 0 serves as a reference / sanity check (questions answered without dialogue history) and τ = 1.0 corresponds to the full context.

Validated use cases (with evidence in the accompanying paper):

• Comparing memory architectures (Full Memory, RAG, Memp, Mem0) across base models (GPT-4o-mini, Gemini-2.5-flash, DeepSeek-V3.2, DeepSeek-V3.2-think) — see the paper.
• Diagnosing temporal retention and forgetting via the Forgetting Measure and Forward-Transfer metrics — see the paper.
• Quantifying per-question inconsistency across checkpoints (i.e., whether a memory system remains correct on the same question over time) via the All-Correct Ratio — see the paper.

Use cases NOT validated:

• Training memory systems (evaluation-only design).
• Memory architectures beyond the four tested (e.g., MemGPT, A-MEM).
• Non-English or cross-cultural dialogue.
• Emotional regulation, empathy, or culturally-grounded reasoning.
• Production deployment readiness in any domain.

Resemblance / utility metrics for the synthetic component (PersonaMem) are reported in the original PersonaMem paper (Jiang et al., 2025); they are not independently re-validated here.

---

Data Limitations

This dataset is a research-only diagnostic benchmark for long-context memory systems and inherits the boundaries of its source data. PersonaMem (Jiang et al., 2025) consists of LLM-synthesized dyadic dialogues, and LongDialQA, released with the DialSim framework (Kim et al., 2024), is constructed from English-language TV scripts (Friends, The Office, The Big Bang Theory). Linguistic and cultural coverage is therefore limited to English dialogue grounded in LLM-generated personas and North American sitcom narratives. Findings on this dataset do not generalize to medical consultations, customer support, group meetings, or non-English / cross-cultural dialogue. The dataset is explicitly NOT recommended for: (a) training memory systems (it was designed for evaluation only and may overfit specific QA formats), (b) certifying deployment readiness in any single application domain — strong benchmark accuracy does not imply real-world safety in healthcare, financial, companionship, or other deployed settings, and (c) commercial deployment without independent domain-specific validation. Four memory architectures are benchmarked (Full Memory, RAG, Memp, Mem0); performance characteristics of other architectures are not directly inferable from these findings.

---

Data Biases

Biases enter this dataset through three layers. (i) PersonaMem (Jiang et al., 2025) is generated by large language models, which are broadly documented in the literature to inherit demographic and cultural skews from their training data; users should consult the original PersonaMem paper for its specific persona distribution and treat its synthetic personas as a non-representative sample of the global population. (ii) LongDialQA (Kim et al., 2024) is derived from US sitcom scripts whose character demographics, humor conventions, and social dynamics reflect predominantly North American narratives; gender, race, age, and cultural representation follow each show's casting choices, not population statistics. Question categories emphasize factual preference tracking and entity identity over emotional nuance or culturally-grounded reasoning, which structurally underrepresents these dimensions. No demographic balancing was performed; downstream users should not interpret accuracy on this benchmark as evidence of equitable behavior across populations.

---

Personal & Sensitive Information

Categories present (inherited from source data):

• Language: English only.
• Geography & Culture: synthetic English-speaking contexts (PersonaMem); North American sitcom settings (LongDialQA).
• Gender, Age, Experience/Seniority: present at the surface level via fictional / synthetic character attributes (e.g., "the user is a graduate student"; "Ross is a paleontologist").
• Political / religious beliefs, Health / medical data, Socio-economic status: NOT systematically collected, but may appear incidentally in free-text dialogue (e.g., a synthetic persona mentioning a dietary preference or hobby).

No real personally identifiable information (PII) is contained. PersonaMem characters are LLM-generated and do not correspond to real individuals. LongDialQA character names (e.g., Ross, Joey) refer to fictional TV characters and are not real persons. No human-subject data was collected by the authors of this benchmark; therefore, no IRB / ethics review was required.

---

Data Social Impact

Positive impact: The dataset enables fine-grained diagnosis of how memory architectures handle preference evolution and long-horizon entity tracking — a gap not well-addressed by existing single-snapshot benchmarks. By exposing temporal forgetting and per-question inconsistency across checkpoints, it supports more rigorous comparison among memory systems and motivates the design of more robust personalized agents.

Risks and mitigations:

• Misuse as training data → The dataset is explicitly designated as evaluation-only and documented in the Limitations section.
• Over-reliance on benchmark accuracy as a proxy for real-world personal-AI safety → Detailed per-checkpoint metrics, including the All-Correct Ratio, are published to discourage interpreting aggregate accuracy as deployment readiness.
• Distributional harm from extending findings to underrepresented populations → The Limitations and Biases sections explicitly list unsupported populations and contexts.
• Reproducibility of long-context evaluation → Processed context files (aggregated from DialSim/LongDialQA) are provided in this repository. These files are redistributed in accordance with the CC BY 4.0 license of the original dataset (Kim et al., 2024) and are intended strictly for non-commercial research purposes. Proper attribution to the original authors is included in the repository documentation.

Access measures: The benchmark is released for non-commercial academic research only. A complete Limitations document accompanies the release. No gated access is required for the derivative annotations.

---

Source Datasets & Licensing

This benchmark is built upon two existing open-source datasets. The derived and processed benchmark is released under the CC-BY-4.0 license, while fully respecting and incorporating the licenses of the original assets.

1. PersonaMem (Jiang et al., 2025)

• URI: https://huggingface.co/datasets/bowen-upenn/PersonaMem
• Original License: MIT License
• Copyright Notice: Copyright (c) 2024 bowen-upenn
• Permission Notice: Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the MIT License conditions.
• Note on Direct Reuse: The PersonaMem context files (shared_contexts_32k.jsonl and shared_contexts_128k.jsonl) are redistributed in this repository as unmodified copies under the original MIT license, with the copyright notice preserved above. They are not claimed as part of the contribution of this benchmark; only the checkpoint-conditioned annotations and evaluation protocol are.

2. LongDialQA (released with the DialSim framework, Kim et al., 2024)

• URI: https://drive.google.com/drive/folders/1MhPlUFWuchVZ5E1NQDWfbT7_RW7ozbuk
• Framework Repository: https://github.com/jiho283/DialSim
• Original License: CC BY 4.0
• Version Used: v1.1
• Attribution: This benchmark adapts data from LongDialQA, released with the DialSim framework (Kim et al., 2024) under CC BY 4.0. The original scripts have been reformatted into long-dialogue contexts for evaluation purposes. Acknowledgment goes to the authors of DialSim for making the multi-party dialogue dataset publicly available. The underlying TV scripts remain copyright of their respective studios; they are used here for non-commercial research purposes only.

---

Citation

If this benchmark is used in research, please cite the accompanying paper:

@inproceedings{anonymous2026comem,
  title={CoMem: A Benchmark for Continual Memory and Dynamic Preference 
         Evolution in Long-Context Agents},
  author={Anonymous Authors},
  booktitle={Submitted to NeurIPS 2026 Datasets and Benchmarks Track},
  year={2026}
}