declare-lab / delta-mem_qwen3_4b-instruct

δ-mem Qwen3-4B Instruct TSW Adapter

This repository contains the δ-mem TSW adapter for Qwen/Qwen3-4B-Instruct-2507 , as presented in the paper δ-mem: Efficient Online Memory for Large Language Models .

δ-mem is a lightweight online memory mechanism that augments a frozen backbone with a compact associative memory state. It projects context into a low-dimensional space and updates a state matrix via delta-rule learning, allowing for efficient long-term memory utilization without full fine-tuning or context extension.

• Paper: https://huggingface.co/papers/2605.12357
• Repository: https://github.com/declare-lab/delta-Mem

Model Details

Usage

To use this adapter, you must first install the $\delta$-mem codebase:

git clone https://github.com/declare-lab/delta-Mem.git
cd delta-Mem
bash scripts/setup_uv_env.sh

Minimal Loading Example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from deltamem.core import HFDeltaMemConfig, attach_delta_mem, load_delta_mem_adapter

base_model = "Qwen/Qwen3-4B-Instruct-2507"
adapter_dir = "declare-lab/delta-mem_qwen3_4b-instruct"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

config = HFDeltaMemConfig.from_pretrained(adapter_dir)
attach_delta_mem(model, config)
load_delta_mem_adapter(model, adapter_dir)
model.eval()

Important Notes

This is not a standard PEFT LoRA adapter and should not be loaded with PeftModel or merged with merge_and_unload() . δ-mem requires its specific runtime memory write/read path. To use this adapter, load the frozen base model, attach the modules, and then load the adapter weights using the provided codebase.

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}, 
}