This repository contains the 1.7B law Memory Decoder released with Memory Decoder at Scale. It is a pretrained parametric long-term memory that can be swapped into a compatible frozen language-model backbone.

This 1.7B memory was trained on the law CPT corpus using retrieval-derived sparse target distributions. The released dataset contains the aligned CPT text, Qwen3-preprocessed data, and KNN distributions. In the paper, the 1.7B Qwen3-vocabulary domain memories are evaluated with frozen Qwen3 Base backbones from 0.6B to 14B.

This checkpoint is a memory component, not a standalone chat- or instruction-tuned model. The memory and backbone must use compatible token IDs and vocabularies.

Model details

Field	Value
Memory size	1.7B class
Architecture/tokenizer family	Qwen3ForCausalLM memory; Qwen3 tokenizer/vocabulary
Domain	Law
Evaluation benchmark	LawBench
Intended backbone	Frozen, tokenizer-compatible Qwen3 Base model
Release contents	Inference weights, configuration, and tokenizer files

Usage

Install the matching environment from LUMIA-Group/MemoryDecoder-at-Scale, then set MODEL_PATH to a compatible frozen backbone and MEMDEC_PATH to this repository:

bash

MODEL_PATH=/path/to/compatible-base-model \
MEMDEC_PATH=Rubin-Wei/MemoryDecoder-Qwen3-1.7B-law \
bash eval/opencompass/scripts/domain/evaluate_lawbench.sh

See the repository documentation and launcher for benchmark-specific options, including interpolation weights and batch settings.

Intended use and limitations

This checkpoint is intended for research and evaluation in the law domain. Its outputs depend on the backbone, prompt, and interpolation settings. Domain specialization does not guarantee factual correctness or safety, and the model may inherit biases and errors from its training sources.

Citation

If you use this checkpoint, please cite:

bibtex

@misc{wei2026memorydecoderscalepretrained,
      title={Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory},
      author={Rubin Wei and Jiaqi Cao and Jiarui Wang and Junming Zhang and Qipeng Guo and Bowen Zhou and Zhouhan Lin},
      year={2026},
      eprint={2607.27919},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.27919},
}