In large language model-based agents, memory serves as a critical capability for achieving personalization by storing and utilizing users' information. Although some previous studies have adopted memory to implement user personalization, they typically focus on preference alignment and simple question-answering. However, in the real world, complex tasks often require multi-hop reasoning on a large amount of user information, which poses significant challenges for current memory approaches. To address this limitation, we propose the multi-hop personalized reasoning task to explore how different memory mechanisms perform in multi-hop reasoning over personalized information. We explicitly define this task and construct a dataset along with a unified evaluation framework. Then, we implement various explicit and implicit memory methods and conduct comprehensive experiments. We evaluate their performance on this task from multiple perspectives and analyze their strengths and weaknesses. Besides, we explore hybrid approaches that combine both paradigms and propose the HybridMem method to address their limitations. We demonstrate the effectiveness of our proposed model through extensive experiments.

• We identify the MPR tasks, highlighting their unique challenges for agent memory compared to previous works.
• We formally define MPR tasks and construct a new dataset with a unified evaluation framework for systematically exploring different memory methods on MPR tasks.
• We conduct comprehensive experiments on explicit, implicit, and hybrid memory approaches, presenting key findings and proposing a new hybrid memory method for long-hop reasoning tasks.

Create an environment with Conda.

conda create -n mpr_env python=3.9

Activate the environment.

conda activate mpr_env

Clone/Dowload the repository source.

git clone https://github.com/nuster1128/MPR.git

Install the packages.

cd MPR
pip install -r requirements.txt

Dowload the MPR Dataset and unzip it.

unzip mpr_dataset.zip

Download the checkpoint of LLM and Sentence Embedding model.

You may refer to Qwen2.5-7B and E5-base-v2.

Please edit the config file config.py to set the proper experiment settings.

Fill out the following paths of model checkpoints and the project.

Qwen_2_5_7B_model_path = '[MODEL_PATH]'
E5_base_v2_model_path = '[MODEL_PATH]'
Root_dir = '[PROJECT_ROOT_PATH]'

Besides, fill out the port and address if using API of LLMs.

Fill out the path of evaluated datasets.

environment_config = {
    'data_title': '[DATA_TITLE]',
    'data_path_dir': '[DATA_DIR_PATH]'
}

Fill out the settings of memory mechanisms.

Fill out the settings of reasoning structures.

Run the program with the commands like the following one.

nohup python -u run_qa.py Naive DenseRAG > logs/Naive-DenseRAG.log 2>&1 &

You may change Naive to other reasoning structures as follows:

• Naive: Naive Reasoning (NR).
• Sequential: Sequential Reasoning (SR).
• Structured: Multi-path Reasoning (MR).
• Decomposition: Decomposition Reasoning (DR).

You may change DenseRAG to other memory mechanisms referred in config.py.

@article{zhang2025explicit,
  title={Explicit vs Implicit Memory: Exploring Multi-hop Complex Reasoning Over Personalized Information},
  author={Zhang, Zeyu and Zhang, Yang and Tan, Haoran and Li, Rui and Chen, Xu},
  journal={arXiv preprint arXiv:2508.13250},
  year={2025}
}

If you have any questions, please feel free to contact us via [email protected].