delta-mem-qwen3-4b-instruct-mlx-adapter

δ-mem Qwen3-4B Instruct MLX Adapter

This repository contains an MLX-converted copy of the public δ-mem TSW adapter for Qwen/Qwen3-4B-Instruct-2507.

δ-mem is an online memory mechanism for frozen LLMs. It adds a compact memory path that writes earlier context into a small online state and reads that state back through attention during later generation. This is an adapter artifact, not a merged base model.

Files

• delta_mem_adapter_mlx.npz: MLX-converted adapter weights.
• delta_mem_config.json: δ-mem adapter configuration copied from the upstream adapter repo.
• conversion_info.json: provenance for this conversion package.

Provenance

• Paper: https://arxiv.org/abs/2605.12357
• Hugging Face paper page: https://huggingface.co/papers/2605.12357
• Upstream Torch adapter: https://huggingface.co/declare-lab/delta-mem_qwen3_4b-instruct
• Upstream adapter commit: c46dc31155608e412d44bf56638d5a6f856f2e7e
• Base model: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507
• My MLX Sidecar repo: https://github.com/elimaine/delta-mem-mlx-sidecar
• My Openclaw integrated plugin: https://clawhub.ai/plugins/@elimaine/openclaw-delta-mem-mlx

Runtime

This adapter is intended for the δ-mem MLX sidecar runtime, not stock mlx_lm.load(). Use it with a compatible MLX Qwen3-4B backbone and a δ-mem-aware attention/runtime implementation.

Example environment shape:

DELTA_MEM_RUNTIME=mlx
DELTA_MEM_MODEL_PATH=mlx-community/Qwen3-4B-Instruct-2507-4bit
DELTA_MEM_ADAPTER_DIR=/path/to/this/adapter
DELTA_MEM_MODEL_ID=delta-mem-qwen3-4b-mlx

Important Notes

• This is not a PEFT LoRA adapter.
• This is not a complete MLX model repo.
• The base model remains frozen; the δ-mem runtime attaches the memory modules and loads these adapter weights.
• Exact hidden-fact recall is not guaranteed. The target behavior is attention-shaped session continuity and measurable memory-heavy benchmark deltas.

Citation

@misc{lei2026deltamemefficientonlinememory,
  title={$\delta$-mem: Efficient Online Memory for Large Language Models},
  author={Jingdi Lei and Di Zhang and Junxian Li and Weida Wang and Kaixuan Fan and Xiang Liu and Qihan Liu and Xiaoteng Ma and Baian Chen and Soujanya Poria},
  year={2026},
  eprint={2605.12357},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2605.12357}
}