MemComposer

These are the trained selectors from MemComposer: Learning to Assemble Memories for LLM Agents via Reinforcement Learning, released as LoRA adapters. For each query, a selector reads the retrieved candidate memories and decides which ones a frozen agent receives, how many, and in what order. It is trained with ExPO, a GRPO-style objective whose reward is the task success of the full agent interaction. Code: github.com/Makr-Xie/memcomposer. Data: Makrrr/MemComposer-data.

Adapters

All adapters are peft LoRA in fp32: r = 64, α = 128, on all linear layers (q/k/v/o_proj, gate/up/down_proj).

subfolder	benchmark	base model (selector and executor)
tau2_qwen_retail	τ²-bench retail	Qwen3-4B-Instruct-2507
tau2_qwen_telecom	τ²-bench telecom	Qwen3-4B-Instruct-2507
tau2_ministral_retail	τ²-bench retail	Ministral-3-8B-Instruct-2512
tau2_ministral_airline	τ²-bench airline	Ministral-3-8B-Instruct-2512
tau2_ministral_telecom	τ²-bench telecom	Ministral-3-8B-Instruct-2512
alfworld_ministral	ALFWorld	Ministral-3-8B-Instruct-2512

Download into the code repository's checkpoints/ directory:

hf download Makrrr/MemComposer --local-dir checkpoints \
    --exclude README.md --exclude .gitattributes

Base models

• Qwen3-4B-Instruct-2507: use Qwen/Qwen3-4B-Instruct-2507 as published.
• Ministral-3-8B-Instruct-2512: use the text-only conversion of mistralai/Ministral-3-8B-Instruct-2512-BF16. Build it with training/tau2_ministral/convert_ministral_text_only.py from the code repository, then apply the tokenizer and RoPE fixes listed in its verl/README.md.

Usage

Attach the adapter to the base model. Do not merge it into bf16 weights: the LoRA update is about 0.1% of the base weights, below bf16 precision, so a bf16 merge erases most of it.

from transformers import AutoModelForCausalLM
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507", dtype="auto")
model = PeftModel.from_pretrained(model, "Makrrr/MemComposer", subfolder="tau2_qwen_retail")

With vLLM, serve the base model with --enable-lora --max-lora-rank 64 and pass an adapter subfolder as a LoRA module.

Each selector expects the prompt format of its training data (see the dataset). It replies with a short <reasoning> block and then an <answer> list of case numbers.

Citation

@misc{xie2026memcomposer,
  title         = {MemComposer: Learning to Assemble Memories for LLM Agents via Reinforcement Learning},
  author        = {Yixuan Xie and Xueqiang Xu and Xinyi Fan and Xinxi Lyu and Tao Feng and Xing Fan and
                   Chenlei Guo and Jie Hao and Shaowen Wang and Jiashuo Sun and Jiawei Han},
  year          = {2026},
  eprint        = {2610.21091},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}