---
license: apache-2.0
library_name: pytorch
language:
  - en
datasets:
  - Salesforce/wikitext
base_model: NousResearch/Llama-2-7b-hf
tags:
  - memory-augmented-llm
  - external-memory
  - engram
  - xmemtransfer
  - arxiv:2608.17050
---

# XMemTransfer LLaMA-2 source memory (15M)

This repository contains a frozen source-memory artifact used in the XMemTransfer experiments.

## Project links

- Website: [XMemTransfer project page](https://olaresearch.github.io/XMemTransfer)
- Paper: [Cross-Model Memory Transfer via Target-Side Reader Adaptation](https://arxiv.org/abs/2608.17050)
- Code: [OLAResearch/XMemTransfer](https://github.com/OLAResearch/XMemTransfer)
- Model repo: [https://huggingface.co/OLAResearchX/xmemtransfer-llama2-source-15m](https://huggingface.co/OLAResearchX/xmemtransfer-llama2-source-15m)
- Collection: [https://huggingface.co/collections/OLAResearchX/xmemtransfer](https://huggingface.co/collections/OLAResearchX/xmemtransfer)

## Files

- `memory.pt`: source memory checkpoint
- `memory_best.pt`: best checkpoint selected during source training
- `memory_config.json`: hash and memory configuration
- `source_adaptor_best.pt`: source-side adaptor checkpoint used with the source memory

## Intended use

This artifact is meant to be paired with target-side adaptors from the XMemTransfer project for cross-model memory-transfer experiments.

## Provenance

- source model family: LLaMA-2 7B
- source corpus: WikiText-2021
- source token budget: 15M

## Usage and limitations

These files are research artifacts for the XMemTransfer implementation, not
standalone Transformers pipeline checkpoints. Follow the
[project code](https://github.com/OLAResearch/XMemTransfer) to construct the
memory reader and load the checkpoint. Results depend on the paired backbone,
source memory, data preprocessing, and configuration released with the model.

## Citation

Please cite the [XMemTransfer paper](https://arxiv.org/abs/2608.17050) when
using these artifacts.