PersonaMem-v2: Towards Personalized Intelligence via Learning Implicit User Personas and Agentic Memory

Bowen Jiang; Yuan Yuan; Maohao Shen; Zhuoqun Hao; Zhangchen Xu; Zichen Chen; Ziyi Liu; Anvesh Rao Vijjini; Jiashu He; Hanchao Yu; Radha Poovendran; Gregory Wornell; Lyle Ungar; Dan Roth; Sihao Chen; Camillo Jose Taylor

Source-linked AI summary

PersonaMem-v2: Towards Personalized Intelligence via Learning Implicit User Personas and Agentic Memory

Bowen Jiang, Yuan Yuan, Maohao Shen, Zhuoqun Hao, Zhangchen Xu, Zichen Chen, Ziyi Liu, Anvesh Rao Vijjini, Jiashu He, Hanchao Yu, Radha Poovendran, Gregory Wornell, Lyle Ungar, Dan Roth, Sihao Chen, Camillo Jose Taylor

arXiv:2512.06688v1Dec 7, 2025cs.CL

TL;DR

PERSONAMEM-V2 addresses the challenge of personalizing responses from long conversations where user preferences are usually implicit. The paper builds a broad benchmark, studies reinforcement fine-tuning, and develops agentic memory; reinforcement-trained and memory-based systems outperform frontier baselines while using substantially fewer tokens. Its scope also includes dataset-quality and privacy boundaries.

Problem
LLMs struggle to infer implicit user personas and preferences from long, noisy conversation histories, although personalization requires responses aligned with individual users.
Method
The paper introduces PERSONAMEM-V2 and uses reinforcement fine-tuning to train personalization reasoning and an agentic memory system that maintains a compact, human-readable user memory.
Results
Agentic memory reaches state-of-the-art implicit personalization while using 16× fewer input tokens, and targeted reinforcement fine-tuning enables a 4B reasoning model to outperform GPT-5.
Takeaways & Limitations
The results identify reasoning as the bottleneck for implicit personalization and present agentic memory as a scalable path toward real-world personalized intelligence.
Takeaways & Limitations
The long-context approach requires appending the entire conversation, becoming inefficient as histories grow; future work includes richer multimodal and interactive memory settings.

Abstract

from arXiv · show

Personalization is one of the next milestones in advancing AI capability and alignment. We introduce PersonaMem-v2, the state-of-the-art dataset for LLM personalization that simulates 1,000 realistic user-chatbot interactions on 300+ scenarios, 20,000+ user preferences, and 128k-token context windows, where most user preferences are implicitly revealed to reflect real-world interactions. Using this data, we investigate how reinforcement fine-tuning enables a model to improve its long-context reasoning capabilities for user understanding and personalization. We also develop a framework for training an agentic memory system, which maintains a single, human-readable memory that grows with each user over time. In our experiments, frontier LLMs still struggle with implicit personalization, achieving only 37-48% accuracy. While they support long context windows, reasoning remains the bottleneck for implicit personalization tasks. Using reinforcement fine-tuning, we successfully train Qwen3-4B to outperforms GPT-5, reaching 53% accuracy in implicit personalization. Moreover, our agentic memory framework achieves state-of-the-art 55% accuracy while using 16x fewer input tokens, relying on a 2k-token memory instead of full 32k conversation histories. These results underscore the impact of our dataset and demonstrate agentic memory as a scalable path toward real-world personalized intelligence.

1 Introduction

PERSONAMEM-V2 targets the challenge of inferring implicit user personas from long, varied interactions and introduces a dataset, reinforcement fine-tuning study, and agentic memory framework for personalization.

Personalized AI must align responses with users’ distinct intentions, contexts, preferences, and emotional states rather than a single correct answer.
Long, noisy conversation histories often reveal preferences only indirectly, requiring models to infer scattered signals while users perform everyday tasks or explore hypothetical prompts.
PERSONAMEM-V2 provides realistic interactions spanning 1,000+ personas, 20,000+ preferences, 300+ topics, and contexts up to 128,000 tokens.
Frontier LLM evaluations expose difficulty interpreting and tracking implicit preferences over long interaction histories, motivating reinforcement fine-tuning for personalization.
The agentic memory framework distills evolving interaction histories into a persistent, human-readable memory and reports state-of-the-art personalization with 16× fewer input tokens.
The paper’s contributions include curating the dataset, benchmarking frontier models, demonstrating reinforcement fine-tuning, and proposing an agentic memory framework.

2 Overview of PERSONAMEM-V2: IMPLICIT PERSONAS

PERSONAMEM-V2 simulates diverse, evolving user interactions in which preferences emerge implicitly across tasks and sessions. Its pipeline combines broad persona coverage with scalable generation and strict validation, including privacy-related scenarios.

2.1 Comprehensive User Personas: The dataset includes 1,000 richly detailed personas spanning demographic, professional, educational, relational, value-based, technological, and conversational attributes.
2.2 User Preferences: It evaluates whether models remember conversationally shared preferences rather than relying on demographic stereotypes, using stereotypical, anti-stereotypical, neutral, health, and therapy-related preferences.
2.3 Dynamic and Ambiguous Preferences: Ambiguous hypothetical or third-person messages and preference changes across sessions test whether models distinguish user behavior and track evolving preferences.
2.4 Cross-Scenario Personalization: Each preference becomes a two-to-six-turn conversation, with cross-scenario cues emerging through everyday tasks such as writing, translation, photo queries, therapy, and medical consultation.
2.4 Cross-Scenario Personalization: Multi-session histories concatenate interactions to 32,000 tokens, while complementary coding and mathematics dialogues extend effective contexts to 128,000 tokens.
2.6 Privacy and Personalization Boundary: The dataset simulates privacy risks involving sensitive information and gives users control over whether such information contributes to personalization.
2.5 Ensuring High Quality of Data: The benchmark includes open-ended and multiple-choice questions generated from preferences, with annotations designed to identify the single personalized option among reasonable alternatives.
2.5 Ensuring High Quality of Data: Only around 30% of generated Q&A pairs survive checks for history dependence, faithful preference representation, unique correctness, and clean formatting.

3 Towards Personalized Intelligence

Personalization requires models to infer implicit preferences from long, noisy conversation histories rather than retrieve explicit facts. The paper trains both long-context reasoning and agentic-memory systems using reinforcement learning with verifiable personalization rewards.

3 Towards Personalized Intelligence: Implicit personalization requires reasoning over subtle preference signals embedded in long user–chatbot histories.Unlike classical question answering, the task involves inferring preferences that users often reveal indirectly.
3 Towards Personalized Intelligence: Long-context reasoning uses the full conversation history but becomes inefficient as the conversation grows.This limitation motivates a more scalable memory mechanism.
3 Towards Personalized Intelligence: GRPO uses MCQ correctness and annotated preference alignment as reward pathways for personalization.These pathways provide verifiable signals for outputs that align with or contradict the user’s persona.
3 Towards Personalized Intelligence: Agentic memory divides histories into chunks and iteratively updates a single capped memory before answering the final query.The memory update uses the current chunk and previous memory, then the final response uses the refined memory.
3 Towards Personalized Intelligence: The memory obeys causality, a Markovian update assumption, and a compact human-readable size constraint.These constraints require storing information before future queries are known while preventing unbounded growth.

4.1 Benchmarking Frontier LLMs in Personalization

Frontier LLMs remain weak at implicit personalization despite long-context capabilities. Their performance varies systematically with preference type, indicating that reasoning and individualized interpretation matter more than simply processing longer histories.

4.1 Benchmarking Frontier LLMs in Personalization: 40-55% accuracy: GPT-5 variants perform only modestly across correlated MCQ and open-ended personalization evaluations.The benchmark evaluates OpenAI models under both task formats, with open-ended responses judged by three independent GPT-5-Chat instances.
4.1 Benchmarking Frontier LLMs in Personalization: No significant accuracy improvement follows shortening contexts from 128k to 32k tokens by removing irrelevant conversations.The result shifts attention from retrieval-style long-context stress tests toward interpreting subtle preference signals.
4.1.1 Frontier LLMs still struggle to infer implicit user preferences: 48.9 ± 5.7% accuracy for stereotypical preferences falls to 41.6 ± 3.3% for neutral and 33.0 ± 5.0% for anti-stereotypical preferences.The pattern suggests reliance on population priors rather than individual behavior.
4.1.1 Frontier LLMs still struggle to infer implicit user preferences: 35.4 ± 10.5% accuracy for dynamic preferences trails 40.1 ± 2.7% for static preferences.This indicates difficulty updating beliefs as user preferences change.
4.1.1 Frontier LLMs still struggle to infer implicit user preferences: 17.5 ± 11.7% accuracy on distinguishing others’ preferences from the user’s own trails 42.1 ± 4.9% for users’ own preferences.Scenario differences are comparatively modest by comparison.

4.2 Training Long-context Reasoning for Personalization

Reinforcement fine-tuning substantially improves Qwen3-4B’s ability to interpret implicit preferences. Mixing MCQ and open-ended supervision provides complementary reward properties, while single-format training produces cross-task degradation.

4.2 Training Long-context Reasoning for Personalization: 53.8% MCQ and 56.0% open-ended accuracy: Qwen3-4B-GRPO surpasses supervised fine-tuning and GPT-5-Chat.GPT-5-Chat reaches 45.6% and 46.2% on the same benchmarks.
4.2.1 Reinforcement learning incentives reasoning toward personalization: Reinforcement fine-tuning enables a 4B-parameter reasoning model to integrate subtle preference signals for personalized responses.The training uses the PERSONAMEM-V2 data and rewards alignment with the current user persona.
4.2.2 Hybrid reward signals unlock better RL toward personalization: 13.1%: MCQ-only training reduces performance on open-ended personalization tasks.Open-ended-only training drops 18.2% on MCQs and 3.7% on open-ended evaluation itself.

4.3 Training Agentic Memory for Personalization

The agentic-memory framework learns a persistent, human-readable representation of each user from chunked conversation histories. It delivers the strongest reported personalization performance while replacing repeated 32k-token processing with a compact 2k-token memory.

4.3 Training Agentic Memory for Personalization: 2k-token memory: the system maintains compactness throughout a 32k-token conversation history.The memory is capped at 2,048 tokens during training.
4.3 Training Agentic Memory for Personalization: 55.2% MCQ and 60.7% open-ended accuracy: agentic memory surpasses direct long-context Qwen3-4B reasoning and GPT-5 models.It achieves the strongest personalization performance across all evaluated models.
4.3.1 Agentic memory delivers state-of-the-art performance with unmatched efficiency: 16× more efficient: compact memory replaces repeatedly processing full 32k-token histories without sacrificing performance.This targets deployment constraints involving latency, cost, and context limits.
4.3.1 Agentic memory delivers state-of-the-art performance with unmatched efficiency: A human-readable memory evolves over time, allowing users to audit, correct, and guide personalization.The framework also suggests offline memory maintenance as a possible deployment strategy.

5 Related Work

Prior personalization work spans retrieval, memory, fine-tuning, and specialized benchmarks, but existing resources incompletely capture dynamic, implicit, real-world preferences. PersonaMem-v2 addresses these gaps by scaling realistic scenarios and evolving preferences.

Methods: Personalization methods include prompt and retrieval approaches, memory-augmented architectures, and preference-based alignment techniques.Examples include user-history retrieval, editable memory graphs, hierarchical memory, RLHF, DPO, and compact personal-preference representations.
Benchmarks: Existing benchmarks cover narrower task settings, cross-user preference simulation, or implicit choices without dynamic preference updates.LaMP uses seven classification and generation tasks; PersonalLLM simulates diverse preference profiles; PrefEval lacks dynamic preference updates.
Benchmark comparison: Table 1 compares PERSONAMEM-V2 with other personalization benchmarks.The supplied passage provides the table caption but no row or column contents.
PERSONAMEM-V2: PERSONAMEM-V2 scales personalization to over 1,000 personas across 300+ scenarios, captures evolving preferences, and emphasizes implicit signals in realistic task-driven interactions.The dataset is presented as a response to limitations in existing personalization data and benchmarks.

6 Conclusion and Future Work

The paper concludes that implicit personalization remains difficult for frontier models, while targeted reinforcement fine-tuning and agentic memory improve performance and efficiency. It identifies richer personalization, interactive memory, privacy control, broader tasks, and real interactions as future directions.

Conclusion: Current frontier LLMs struggle with implicit preferences, while reasoning—not longer context handling or memorization—drives personalization success.Performance also varies systematically across preference types, suggesting reliance on population priors rather than individual behavior.
Future Work: Future work includes richer multimodal personalization, more structured interactive memories, additional utility tasks, user-controlled privacy boundaries, and more realistic real-interaction training data.These directions are presented as opportunities for extending personalized intelligence and agentic memory.