Evoking User Memory: Personalizing LLM via Recollection-Familiarity Adaptive Retrieval

This repository contains the code and resources for the ICLR 2026 paper titled "Evoking User Memory: Personalizing LLM via Recollection-Familiarity Adaptive Retrieval".

The origin code of submission can be found here: https://openreview.net/attachment?id=f7p0F2X6XN&name=supplementary_material.

🧠 RF-Mem

RF-Mem (Recollection–Familiarity Memory) is an adaptive retrieval framework for Large Language Models, grounded in the dual-process theory of human cognition.
In cognitive science, memory is often described as a combination of two systems:

• Familiarity – fast, effortless recognition that something has been seen before.
• Recollection – slower, deliberate reconstruction of episodic details such as time, place, and context.

This theory motivates RF-Mem: instead of treating retrieval as either one-shot search or full-context prompting, RF-Mem adaptively balances the two.

📊 Framework Overview

The RF-Mem framework first measures retrieval uncertainty using mean similarity and entropy.

• If ths system feel familiarity, it selects the Familiarity path for direct retrieval.
• If this system feel unfamiliar, it switches to the Recollection path, where memory items are clustered and iteratively expanded with query reformulation.

This dual-path design enables scalable and efficient personalized memory retrieval, improving both accuracy and efficiency across long-term memory benchmarks such as PersonaBench, PersonaMem, and LongMemEval.

---

📂 Project Structure

RF_mem
├──LongMemEval/.                # LongMemEval experiment data and scripts
│   ├──  assets/                # Figures of this paper
│   ├──  data/                  # Shared data directory
│   ├──  LICENSE
│   ├──  README.md              # Project description
│   ├──  requirements-full.txt  # Full dependencies
│   ├──  requirements-lite.txt  # Lightweight dependencies
│   ├──  run.sh                 # Quick start script
│
├── personabench_data/          # PersonaBench experiment data and scripts
│   ├── eval_data/              # Evaluation data
│   ├── llm_model/              # LLM model wrapper
│   ├── retri_mdoel/            # Retrieval model
│   ├── eval.py                 # PersonaBench evaluation entry
│   ├── main_batch.py           # Batch experiment entry
│   ├── run.sh                  # Quick run script
│   └── utils.py                # Utility functions
│
└── personamem_data/            # PersonaMem experiment data and scripts
    ├── apitokens/              # API key configuration
    ├── data/                   # Data directory
    ├── llm_model/              # LLM model wrapper
    ├── retri_mdoel/            # Retrieval model
    ├── main_batch.py           # PersonaMem batch experiment entry
    ├── run.sh                  # Quick run script
    └── utils.py                # Utility functions

---

🔧 Installation

It is recommended to use Python 3.10+.

pip install -r requirements.txt

---

🚀 Usage

1. Run LongMemEval experiments

   cd LongMemEval
   bash run.sh

2. Run PersonaBench experiments

   cd personabench_data
   bash run.sh

3. Run PersonaMem experiments

   cd personamem_data
   bash run.sh

---

📜 License

This project is licensed under the MIT License.

---

✨ Acknowledgements

• Authors of LongMemEval, PersonaBench & PersonaMem datasets
• OpenAI / HuggingFace model community