[ACL 2026] MemSearcher

📃 Paper | 🤗 Models

MemSearcher is a search agent that keeps a compact, iteratively-updated memory instead of the full interaction history, trained end-to-end with multi-context GRPO.

---

Environment Setup

conda create -n memsearcher python=3.10 -y
conda activate memsearcher
pip install --upgrade pip
pip install --no-deps -r requirements.txt
pip install -e . --no-deps
pip install flash-attn --no-build-isolation
conda install -c pytorch -c nvidia faiss-gpu=1.8.0

The memsearcher environment is used for training (vLLM rollout), and the retriever.

Model serving (evaluation) uses SGLang, which requirements.txt does not include — install it in a separate environment:

conda create -n memsearcher-sglang python=3.10 -y && conda activate memsearcher-sglang
pip install "sglang[all]==0.4.9"

---

MemSearcher Weights

Please check out our Model Zoo for all public MemSearcher checkpoints.

---

Evaluation

1. Download the corpus + prebuilt index (wiki-18 + E5 index):

bash scripts/serving/download_corpus.sh

2. Start the retriever server:

cd scripts/serving
python retriever_serving.py --config retriever_config.yaml --num_retriever 1 --port 8000

3. Serve the model with SGLang (in the memsearcher-sglang env):

conda activate memsearcher-sglang
MODEL_PATH=yuanqianhao/MemSearcher-3B bash launch_server.sh

4. Run the evaluation:

SAVE_NOTE=MemSearcher-3B \
DATA_DIR=./data \
SGL_REMOTE_URL=http://127.0.0.1:80 RETRIEVER_URL=http://127.0.0.1:8000/search \
DATASETS="nq triviaqa popqa hotpotqa 2wikimultihopqa musique bamboogle" \
bash eval.sh

---

Training

Train MemSearcher from a Qwen2.5-Instruct model with multi-context GRPO.

1. Download the corpus + prebuilt index (same artifacts as evaluation):

bash scripts/serving/download_corpus.sh

2. Start the retriever server — the rollout searches it at SEARCH_URL:

cd scripts/serving
python retriever_serving.py --config retriever_config.yaml --num_retriever 1 --port 8000

3. Prepare the training data — the NQ + HotpotQA training split, re-wrapped into MemSearcher's RL format:

python data/prepare_train_data.py --data_sources nq,hotpotqa \
  --output data/nq+hotpotqa_train_converted.parquet

4. Launch training:

MODEL_PATH=Qwen/Qwen2.5-3B-Instruct \
SAVE_PATH=checkpoints/MemSearcher-3B \
SEARCH_URL=http://127.0.0.1:8000 \
bash train.sh

5. Merge the FSDP shards into HuggingFace weights, then evaluate via the Evaluation section:

bash model_merge.sh checkpoints/MemSearcher-3B

---

Citation

If you find MemSearcher useful for your research and applications, please cite using this BibTeX:

@article{yuan2025memsearcher,
  title={MemSearcher: Training LLMs to Reason, Search and Manage Memory via End-to-End Reinforcement Learning},
  author={Yuan, Qianhao and Lou, Jie and Li, Zichao and Chen, Jiawei and Lu, Yaojie and Lin, Hongyu and Sun, Le and Zhang, Debing and Han, Xianpei},
  journal={arXiv preprint arXiv:2511.02805},
  year={2025}
}