MLP Memory: A Retriever-Pretrained Memory for Large Language Models

MLP Memory: A Retriever-Pretrained Memory for Large Language Models

ICLR 2026

📖 Overview

MLP Memory introduces a retriever-pretrained parametric memory that bridges the gap between retrieval-augmented generation (RAG) and parametric fine-tuning. Instead of explicitly fetching documents, it learns to internalize retrieval patterns by pretraining an MLP to mimic kNN retrievers over the entire pretraining corpus. It offers several key advantages.

• ⚙️ End-to-End Differentiable — Unlike non-parametric retrievers, MLP Memory is fully parameterized and supports gradient flow, enabling joint optimization with the base model.
• 💾 Highly Compressed Knowledge — Compresses massive retrieval stores (e.g., 40 TB for 5 B tokens) into a compact 1 B-parameter MLP (~4 GB) while improving overall performance.
• ⚡ Efficient Inference — Eliminates retrieval overhead, achieving faster inference than RAG and kNN-LM, with constant speed regardless of corpus size.
• 🧠 Long-Term Memory — Functions as a durable repository capturing the full pretraining corpus, extending beyond short-term context memory.

Unlike retrieval-augmented methods that suffer from high latency and shallow integration, or parametric fine-tuning approaches that risk catastrophic forgetting, MLP Memory achieves both efficiency and effectiveness, establishing a new paradigm for retrieval-inspired knowledge augmentation in large language models.

🚀 Quick Start

🔧 Environment Setup

We run on CUDA 12.1 with faiss-gpu 1.8.0 as the core dependency. To quickly set up the environment, simply create it from the provided configuration file:

conda env create -f environment.yml
conda activate mlpmemory

🤗 Models and Datasets

Models

• GPT2-large-finetuned-wikitext103: The base model enhanced by MLP Memory, fine-tuned on the WikiText-103 dataset.
• MLPMemory-gpt2-large: The MLP Memory module trained for GPT2-large-finetuned-wikitext103.
• MLPMemory-Mistral-wikiepdia: The MLP Memory trained on the English Wikipedia (Dec 2021) corpus to augment the Mistral-7B-v0.3 model.

Datasets

• Preprocessed English Wikipedia (Dec 2021): The English Wikipedia dataset (Dec 2021) preprocessed using the Mistral tokenizer.
• kNN-Targets-wikipedia-mistral: The kNN target dataset generated from the Wikipedia corpus for the Mistral model, used directly for MLP Memory training.

📊 Evaluate and Use MLP Memory

We provide the checkpoint of gpt2-large MLP Memory used in our experiments 🤗MLPMemory-gpt2-large. Simply download this checkpoint and 🤗wikitext-103 dataset from huggingface and run the following scripts:

📝 Data Preprocessing

# scripts/preprocess_dataset.sh
TOKENIZER="/path/to/tokenizer(model)/directory"
OUTPUT_DIR=./dataset/wikitext-gpt2

python utils/preprocess_dataset.py \
    --dataset_name /path/to/wikitext \
    --dataset_config_name wikitext-103-raw-v1 \
    --tokenizer_path ${TOKENIZER} \
    --output_dir ${OUTPUT_DIR} \
    --num_proc 1

📈 Evaluate Base Model

# scripts/evaluate_base.sh
DATASET=/path/to/dataset
MODEL=/path/to/base/model
OUTPUT_DIR=tmp/

NCCL_P2P_DISABLE=1 NCCL_IB_DISABLE=1 CUDA_VISIBLE_DEVICES=0 python \
    -m train_base \
    --model_name_or_path ${MODEL} \
    --dataset_name ${DATASET} \
    --per_device_eval_batch_size 8 \
    --do_eval \
    --eval_subset test \
    --output_dir ${OUTPUT_DIR} \
    --report_to none

🎯 Evaluate with MLP Memory

# scripts/evaluate_joint.sh
DATASET=/path/to/dataset
MODEL=/path/to/base/model
KNN_PATH=/path/to/mlp/memory
OUTPUT_DIR=tmp/

python -m evaluate_joint \
    --do_test \
    --model_name_or_path ${MODEL} \
    --dataset_name ${DATASET} \
    --dataset_split_name test \
    --per_device_eval_batch_size 8 \
    --output_dir ${OUTPUT_DIR} \
    --knn_temp 1 \
    --lmbda 0.25 \
    --knn_generator_path ${KNN_PATH} \
    --report_to none

🏆 Performance Results on WikiText-103

Model	#Params	PPL
GPT2-large-vanilla	774M	15.80
GPT2-large-finetuned	774M	10.42
GPT2-xl-finetuned	1.5B	10.16
GPT2-large-finetuned + MLPMem	774M + 774M	9.58

💡 Generation Example

# demo/generation_example.py
python -m demo.generation_example

📊 Generation Results Comparison:

Model	Generated Continuation
Base Model	"...who sings i can't take my eyes off of you?? The answer is: Andy Williams..."
+MLP Memory	"...who sings i can't take my eyes off of you?? The answer is: Frankie Valli. ;)..."

🛠️ Training MLP Memory

📁 Repository Structure

Our codebase is organized as follows to facilitate both training and evaluation:

MLPMemory/
├── demo
│   └── generation_example.py      # Generation demonstration
├── downstream
│   ├── eval_qa.py                 # QA tasks evaluation
│   └── eval_qa.sh                 # Evaluate QA tasks
├── knn_utils
│   ├── build_index.py             # build FAISS index for efficient search
│   ├── saveEmbedMulti.py          # Save embeddings with multi-GPU support
│   └── saveKNNMulti.py            # Search and save KNN distributions
├── models
│   ├── __init__.py
│   ├── mlpGPT2.py                 # MLP Memory for GPT2
│   ├── mlpLlama.py                # MLP Memory for Llama
│   ├── mlpMemory.py               # Class for MLP Memory Generation & Training
│   └── mlpMistral.py              # MLP Memory for Mistral
├── scripts
│   ├── evaluate_base.sh           # Evaluate base model
│   ├── evaluate_joint.sh          # Evaluate with MLP Memory
│   ├── preprocess_dataset.sh      # Preprocess datasets
│   ├── save_pipeline.sh           # Complete KNN signal pipeline
│   ├── train_mlpmem_offline.sh    # Train MLP Memory offline
│   └── train_mlpmem_online.sh     # Train MLP Memory online
├── utils
│   ├── cal_loss.py                # Loss calculation utilities
│   └── preprocess_dataset.py      # Dataset preprocessing
├── environment.yml                # Environment configuration
├── evaluate_joint.py              # Joint evaluation interface
├── train_base.py                  # Base model training/evaluation
├── train_mlpmem_offline.py        # MLP Memory training offline
└── train_mlpmem_online.py         # MLP Memory training online

🔄 Training Pipeline

1️⃣ Preprocess Dataset

Tokenize and group text for efficient processing:

bash scripts/preprocess_dataset.sh

2️⃣ Build KNN Training Signals

Three-step process for creating supervision signals:

• Save Embeddings

Extract and save hidden representations from the pretrained model:

accelerate launch \
    --config_file ${ACCELERATE_CONFIG} \
    -m train_base \
    --model_name_or_path ${MODEL_TO_SAVE} \
    --dataset_name ${DATASET} \
    --do_eval --eval_subset ${SUBSET} \
    --per_device_eval_batch_size ${BATCH_SIZE_EVAL} \
    --output_dir ${OUTPUT_DIR} \
    --dstore_dir ${DSTORE_DIR} \
    --save_knnlm_dstore \
    --report_to none

• Build IVFPQ Index

Create an efficient index for fast nearest neighbor search:

python -m knn_utils.build_index \
    --dstore_path ${DSTORE_PATH} \
    --num_keys_to_add_at_a_time ${NUM_KEYS_TO_ADD} \
    --ncentroids ${NCENTROIDS} \
    --code_size ${CODE_SIZE} \
    --probe ${PROBE}

• Search KNN Distributions

Generate KNN probability distributions as training signals:

accelerate launch \
    --config_file ${ACCELERATE_CONFIG} \
    -m knn_utils.saveKNNMulti \
    --model_path ${MODEL_TO_SAVE} \
    --dstore_path ${DSTORE_PATH} \
    --val_path ${VAL_PATH} \
    --index_path ${INDEX_PATH} \
    --output_path ${OUTPUT_PATH} \
    --k ${K} \
    --knn_temp ${KNN_TEMP} \
    --probe ${PROBE} \
    --batch_size ${BATCH_SIZE_KNN} \
    --ignore_first True \
    --knn_gpu

The complete pipeline is available in:

bash scripts/save_pipeline.sh

3️⃣ Start Training

Depending on your storage and compute setup, MLP Memory supports two training modes:

🧱 Offline Training

If you have sufficient storage, we recommend offline training. In this mode, all base model embeddings are precomputed and stored before training begins. This approach requires more disk space but enables faster training since embeddings are directly loaded from disk.

To prepare the embeddings, first run the Save Embeddings step.

Then start the offline training process:

# scripts/train_mlpmem_offline.sh
bash train_mlpmem_offline.sh

🌐 Online Training

If storage is limited, you can opt for online training. In this mode, embeddings are computed dynamically during training rather than pre-stored, which greatly reduces disk usage but incurs a small runtime overhead.

# scripts/train_mlpmem_online.sh
bash train_mlpmem_online.sh

💡 Downstream Evaluation

To eval base model, remove --use_neural_knn and run:

# downstream/eval_qa.sh
bash downstream/eval_qa.sh

To eval base model + MLP Memory, run:

# downstream/eval_qa.sh
bash downstream/eval_qa.sh

🙏 Acknowledgments

This implementation is inspired by the excellent work in knn-transformers and MemoryDecoder. We are grateful for their pioneering contributions to retrieval-augmented language modeling. This work is Supported by Shanghai General Al Foundation Models Program (Grant No.2025SHZDZX025G09)- Technical collaboration: Shanghai Artificial Intelligence Laboratory.

📧 Contact

For questions and discussions, feel free to email: weirubinn@gmail.com

📚 Citation

If you find MLP Memory helpful in your research, please consider citing:

@article{wei2025mlp,
  title={MLP Memory: A Retriever-Pretrained Memory for Large Language Models},
  author={Wei, Rubin and Cao, Jiaqi and Wang, Jiarui and Kai, Jushi and Guo, Qipeng and Zhou, Bowen and Lin, Zhouhan},
  journal={arXiv preprint arXiv:2508.01832},
  year={2025}
}