---
library_name: transformers
license: apache-2.0
tags:
- memory-decoder
- parametric-memory
- long-term-memory
- memorydecoder-at-scale
- text-generation
- pythia
---
# MemoryDecoder-Pythia-1.4B-general

## Resources

- [Project Page: Memory Decoder at Scale](https://rubin-wei.github.io/memory-decoder-at-scale/)
- [GitHub Repository: LUMIA-Group/MemoryDecoder-at-Scale](https://github.com/LUMIA-Group/MemoryDecoder-at-Scale)
- [Paper: Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory](https://arxiv.org/abs/2607.27919)
- [Hugging Face Collection: MemoryDecoder-at-Scale](https://huggingface.co/collections/Rubin-Wei/memorydecoder-at-scale)

This repository contains the **1.4B general Memory Decoder** released with
[Memory Decoder at Scale](https://arxiv.org/abs/2607.27919). It is a pretrained parametric long-term
memory designed to be paired with a frozen Pythia-family language model. The
general-memory suite was pretrained on 300B tokens and scales memory capacity
independently from the backbone.

This checkpoint is a memory component, not a chat- or instruction-tuned model.
Use it through the Memory Decoder integration in the released codebase rather
than treating it as a standalone assistant.

## Model details

| Field | Value |
|---|---|
| Memory size | approximately 1.4B parameters |
| Architecture/tokenizer family | GPT-NeoX / Pythia |
| Memory scope | General |
| Training scale | 300B-token general-memory pretraining suite |
| Intended backbone | Frozen, tokenizer-compatible Pythia model |
| Release contents | Inference weights, configuration, and tokenizer files |

## Usage

Install the environments from
[LUMIA-Group/MemoryDecoder-at-Scale](https://github.com/LUMIA-Group/MemoryDecoder-at-Scale), then use the
`hf-memdec` adapter. For example:

```bash
cd eval/lm-evaluation-harness

BACKBONE=EleutherAI/pythia-410m-deduped
MEMORY=Rubin-Wei/MemoryDecoder-Pythia-1.4B-general

lm-eval \
  --model hf-memdec \
  --model_args pretrained=$BACKBONE,memdec_path=$MEMORY \
  --tasks arc_easy,piqa,mmlu \
  --batch_size 1
```

The adapter provides the task-specific interpolation settings used for the
released Pythia memories; an explicit `lmbda` can be supplied as an override.

## Intended use and limitations

This checkpoint is intended for research on parametric memory, memory scaling,
and evaluation with frozen language-model backbones. Its behavior depends on
the selected backbone and interpolation settings. It may reproduce biases or
errors present in its training data and should not be treated as an authoritative
knowledge source.

## 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},
}
```