ahmedehabb/Memory-R2-answer-agent

Memory-R2 7B — Answer Agent

The trained answer agent (sft_cont_step55) from Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents (arXiv:2605.21768).

It is a Qwen2.5-7B-Instruct model trained with an SFT warm-start followed by an RL continuation (answer-F1 reward). Given a question and a memory store, it generates the final answer.

This model only answers questions — it does not manage memory. It is meant to be paired with the Memory-R2 memory manager, which reads the running conversation and maintains the memory store this model answers from. It is optional and swappable: the memory manager was evaluated against several different answer agents in the paper (untrained Qwen-7B, GPT-OSS-120B, this one) — any instruction-tuned LLM can play this role, and using a different one has no effect on how memory is maintained.

Headline results (tab:main)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

answer_agent = AutoModelForCausalLM.from_pretrained("ahmedehabb/Memory-R2-answer-agent", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("ahmedehabb/Memory-R2-answer-agent")

Full inference code and the memory-store protocol are in the project repository (see the paper for the official release).

Training

• —Base model: Qwen/Qwen2.5-7B-Instruct
• —SFT warm-start followed by an RL continuation (answer-F1 reward) against the memory manager's rollouts
• —Judge for reward/logging during training: GPT-OSS-120B

Citation

@misc{yan2026memoryr2faircreditassignment,
      title={Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents},
      author={Sikuan Yan and Ahmed Bahloul and Ercong Nie and Susanna Schwarzmann and Riccardo Trivisonno and Volker Tresp and Yunpu Ma},
      year={2026},
      eprint={2605.21768},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.21768},
}