This is the official repository of PersonaMem-v2: Towards Personalized Intelligence via Learning Implicit User Personas and Agentic Memory

🎉 Our dataset has been downloaded over 30,000 times on Hugging Face. PersonaMem-v3 is already inside our lab and will be released soon. Thank you for your support!

PersonaMem-v2 is the state-of-the-art LLM personalization benchmark, covering 1000 comprehensive user personas and 20,000+ preferences across 300+ scenarios. It especially focuses on realistic cases where user preferences are implicitly revealed through long-context conversations rather than stated explicitly. We simulate realistic long-form chat histories to test how well LLMs and agentic memory systems can infer these signals and deliver personalized responses over time.

---

Installation

Environment

We use Python 3.10 with CUDA 12.6. A Dockerfile is provided:

# Build
docker build -t persona_mem .

# Run with all GPUs
docker run -it --gpus all -v /path/to/PersonaMem-v2:/workspace persona_mem /bin/bash

API Keys

cp .env.example .env

Then fill in your credentials in .env by choosing one of the following:

OpenAI

OPENAI_API_KEY=your_openai_api_key_here
OPENAI_MODEL=gpt-5-chat
OPENAI_MODEL_EMBED=text-embedding-3-large

Microsoft Azure

AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
AZURE_OPENAI_KEY=your_azure_openai_key_here
AZURE_OPENAI_DEPLOYMENT_NAME=gpt-5-chat
AZURE_OPENAI_API_VERSION=your_api_version
AZURE_OPENAI_DEPLOYMENT_NAME_EMBED=text-embedding-3-large
AZURE_OPENAI_API_VERSION_EMBED=your_embed_api_version

Data

Download the data from HuggingFace and place it at data/.

---

Inference and Evaluation

To run inference over the benchmark data/benchmark/multimodal/benchmark.csv on frontier LLMs, run

bash scripts/inference_scripts/run_gpt5_chat.sh

Inference scripts for other models are in scripts/inference_scripts/.

---

Training

First, download the initial model checkpoint:

bash verl_custom/scripts/download_model.sh

Train a model using vanilla GRPO over long context with the verl framework:

Our modified training code is kept under verl_custom/ separately from the original verl for clarity.

Prepare the data with the format needed for training:

python verl_custom/data_preprocess_rft.py
python verl_custom/data_preprocess_sft.py

Optionally, cold-start the base model with SFT before GRPO:

bash verl_custom/scripts/run_qwen3_4b_sft.sh

Then run training with GRPO:

bash verl_custom/scripts/run_qwen3_4b_grpo.sh

Run inference with a trained checkpoint:

bash verl_custom/scripts/run_qwen3_4b_inference.sh

See verl_custom/ for configuration details and model options.

Train a model using Agentic Memory with the verl framework:

We adapt the original MemAgent pipeline, but shift its focus from long-context knowledge QA to scalable personalization and user memory that grows with each user over time. Our modifications also remove the original dependency on future user queries, which violates causality principles in a realistic personalization setting.

First, prepare the data with the format needed for training:

python MemAgent/data/data_preprocess.py

Then run training with GRPO:

bash MemAgent/run_qwen3_4b_grpo.sh

Run inference with a trained checkpoint:

bash MemAgent/run_qwen3_4b_inference.sh

See MemAgent/ for configuration details and model options.

---

Data Generation

To regenerate the benchmark from scratch, run the full pipeline:

bash scripts/data_gen_scripts/run_generate_all.sh

This runs five sequential steps: image embedding setup, conversation generation, Q&A generation, chat history building (32k and 128k), and benchmark CSV preparation. Individual steps are available as separate scripts in scripts/data_gen_scripts/.

Key parameters in the script:

• --num_persona: number of personas to generate (default: 1000)
• --rate_limit_per_min: API rate limit
• --parallel: enable parallel processing

Run PYTHONPATH=. python data_generation/main.py --help for all options.

See data_generation/ for more configuration details.

Citations

If you find our work helpful, please consider cite them. Thank you!

@article{jiang2026personamem,
  title={PersonaMem-v3: Toward Omni-Platform Personal Intelligence for Holistic User Understanding, Recommendation, and Agentic Tasks},
  author={Jiang, Bowen and Yuan, Yuan and Hao, Zhuoqun and Liu, Yuchen and Shen, Maohao and Chen, Sihao and Wornell, Gregory and Callison-Burch, Chris and Ungar, Lyle and Roth, Dan and others},
  journal={arXiv preprint arXiv:2608.21381},
  year={2026}
}

@article{jiang2025personamem2,
  title={PersonaMem-v2: Towards Personalized Intelligence via Learning Implicit User Personas and Agentic Memory},
  author={Jiang, Bowen and Yuan, Yuan and Shen, Maohao and Hao, Zhuoqun and Xu, Zhangchen and Chen, Zichen and Liu, Ziyi and Vijjini, Anvesh Rao and He, Jiashu and Yu, Hanchao and Poovendran, Radha and Wornell, Gregory and Ungar, Lyle and Roth, Dan and Chen, Sihao and Taylor, Camillo Jose},
  journal={arXiv preprint arXiv:2512.06688},
  year={2025}
}

@article{jiang2025know,
  title={Know Me, Respond to Me: Benchmarking LLMs for Dynamic User Profiling and Personalized Responses at Scale},
  author={Jiang, Bowen and Hao, Zhuoqun and Cho, Young-Min and Li, Bryan and Yuan, Yuan and Chen, Sihao and Ungar, Lyle and Taylor, Camillo J and Roth, Dan},
  journal={arXiv preprint arXiv:2504.14225},
  year={2025}
}