| MobiAgent Paper | MobiMem Paper | Huggingface | App |

English | 中文

MobiAgent is a powerful and customizable mobile agent system including:

• An agent model family: MobiMind
• An agent acceleration framework: AgentRR
• An agent benchmark: MobiFlow

System Architecture:

• [2025.12.26] 📱 Pure on-device inference on smartphones is now supported! See phone_runner/README.md to get started.
• [2025.12.25] 🛠️ We've released unified GUI agent runner supporting one-click config of multiple models (MobiAgent, UI-TARS, AutoGLM, Qwen-VL, Gemini, etc.). See Unify Runner README to get started.
• [2025.12.08] 🔥 We've released MobiMind-Reasoning-4B and its quantized version MobiMind-Reasoning-4B-AWQ.
• [2025.11.03] Added multi-task execution support. See Multi-task README for details.
• [2025.11.03] Introduced a user profile memory system, enabled via --user_profile on. See User Profile README for details.

• [2025.8.30] We've open-sourced MobiAgent!

Mobile App Demo:

AgentRR Demo (Left: first task; Right: subsequent task)

Multi Task Demo

task: 在小红书查找2025年性价比最高的单反相机推荐，然后在淘宝搜索该相机，并将淘宝中的相机品牌、名称和价格通过微信发送给小赵。

• agent_rr/ - Agent Record & Replay framework
• collect/ - Data collection, annotation, processing and export tools
• runner/ - Agent executor that connects to phone via ADB, executes tasks, and records execution traces
• MobiFlow/ - Agent evaluation benchmark based on milestone DAG
• app/ - MobiAgent Android app
• deployment/ - Service deployment for MobiAgent mobile application

If you would like to try MobiAgent directly with our APP, please download it in Download Link and enjoy yourself!

If you would like to try MobiAgent with python scripts which leverage Android Debug Bridge (ADB) to control your phone, please follow these steps:

Create virtual environment, e.g., using conda:

conda create -n MobiMind python=3.10
conda activate MobiMind

Simplest environment setup (in case you want to run the agent runner alone):

# Install simplest dependencies
pip install -r requirements_simple.txt

Full environment setup (in case you want to run the full pipeline):

pip install -r requirements.txt

# Download OmniParser model weights
for f in icon_detect/{train_args.yaml,model.pt,model.yaml} ; do huggingface-cli download microsoft/OmniParser-v2.0 "$f" --local-dir weights; done

# Download embedding model utils
huggingface-cli download BAAI/bge-small-zh --local-dir ./utils/experience/BAAI/bge-small-zh

# Install OCR utils (optional)
sudo apt install tesseract-ocr tesseract-ocr-chi-sim

# If you need GPU acceleration for OCR, install paddlepaddle-gpu according to your CUDA version
# For details, refer to https://www.paddlepaddle.org.cn/install/quick, CUDA 11.8 for example:
python -m pip install paddlepaddle-gpu>=3.1.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/

• Download and install ADBKeyboard on your Android device
• Enable Developer Options on your Android device and allow USB debugging
• Connect your phone to the computer using a USB cable

After downloading the model checkpoints, use vLLM to deploy model inference services:

download urls:

• MobiMind-1.5-4B(fastest&experimental, support e2e):
  • huggingface
  • modelscope
• MobiMind-Reasoning-4B(stable):
  • huggingface
  • modelscope

vllm serve MobiMind-Reasoning-4B --port <decider/grounder port>
vllm serve Qwen/Qwen3-4B-Instruct --port <planner port>

MobiAgent supports three types of memory systems to enhance agent performance:

User preference memory system (Mem0) provides personalized context for planning. To enable it, set up the backend storage:

Milvus (Vector Database) - Required for vector search:

# Download the installation script
curl -sfL https://raw.githubusercontent.com/milvus-io/milvus/master/scripts/standalone_embed.sh -o standalone_embed.sh
# Start the Docker container
bash standalone_embed.sh start

Add to your .env file:

MILVUS_URL=http://localhost:19530
EMBEDDING_MODEL=BAAI/bge-small-zh
EMBEDDING_MODEL_DIMS=384
OPENAI_API_KEY=your_key_here
OPENAI_BASE_URL=your_llm_endpoint_here

Neo4j (GraphRAG) - Optional for graph-based retrieval:

docker run -d --name neo4j \
  -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/testpassword \
  neo4j:5.23.0

Add to your .env file:

NEO4J_URL=neo4j://localhost:7687
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=testpassword

For detailed configuration, see runner README.

Experience memory enables the planner to retrieve and use similar past task execution experiences. Enable it by adding the --use_experience flag when launching the agent runner.

Action memory (AgentRR) caches and reuses successful action sequences to accelerate task execution. For ActTree reproduction and evaluation, see AgentRR README (ActTree). For ActChain (experience-based action memory) which is being integrated with Agent Runner as an experimental feature, see #49.

Write the list of tasks that you would like to test in runner/mobiagent/task.json, then launch agent runner:

Basic launch:

python -m runner.mobiagent.mobiagent \
  --service_ip <Service IP> \
  --decider_port <Decider Service Port> \
  --grounder_port <Grounder Service Port> # Same as Decider Port, be ignored when using --e2e flag \
  --planner_port <Planner Service Port>

With user profile memory:

python -m runner.mobiagent.mobiagent \
  --service_ip <Service IP> \
  --decider_port <Decider Service Port> \
  --grounder_port <Grounder Service Port> # Same as Decider Port, be ignored when using --e2e flag \
  --planner_port <Planner Service Port> \
  --user_profile on \
  --use_graphrag off  # Use 'on' for GraphRAG (Neo4j), 'off' for vector search (Milvus)

Common parameters:

• --service_ip: Service IP (default: localhost)
• --decider_port: Decider service port (default: 8000)
• --grounder_port: Grounder service port (default: 8001)
• --planner_port: Planner service port (default: 8002)
• --e2e: Use end-to-end model and elinimates grounder calls to accelerate execution (default: false)
• --device: Device type, Android or Harmony (default: Android)
• --user_profile: Enable user profile memory, on or off (default: off)
• --use_graphrag: Use GraphRAG (Neo4j) for retrieval, on or off (default: off)
• --use_experience: Enable experience-based task rewriting (default: False)
• --data_dir: Directory to save result data (default: runner/mobiagent/data/)
• --task_file: Path to the task list file (default: runner/mobiagent/task.json)

The runner automatically controls the device and invoke agent models to complete the pre-defined tasks.

Important: If you deploy MobiMind-Reasoning-4B model inference, set both decider/grounder ports to <decider/grounder port>.

For all available parameters, see runner README.

Multi-task Execution:

For complex tasks that require interaction with multiple applications, use the multi-task executor:

python -m runner.mobiagent.multi_task.mobiagent_refactored \
  --service_ip <Service IP> \
  --decider_port <Decider Service Port> \
  --grounder_port <Grounder Service Port> \
  --planner_port <Planner Service Port> \
  --task "Your multi-step task description"

For detailed configuration, multi-screenshot support, OCR settings, and experience memory integration, see Multi-task README.

For detailed usage instructions, see the README.md files in each sub-module directory.

If you find MobiAgent useful in your research, please feel free to cite our paper:

@misc{zhang2025mobiagentsystematicframeworkcustomizable,
  title={MobiAgent: A Systematic Framework for Customizable Mobile Agents}, 
  author={Cheng Zhang and Erhu Feng and Xi Zhao and Yisheng Zhao and Wangbo Gong and Jiahui Sun and Dong Du and Zhichao Hua and Yubin Xia and Haibo Chen},
  year={2025},
  eprint={2509.00531},
  archivePrefix={arXiv},
  primaryClass={cs.MA},
  url={https://arxiv.org/abs/2509.00531}, 
}
@misc{liu2025trainingenablingselfevolutionagents,
  title={Beyond Training: Enabling Self-Evolution of Agents with MOBIMEM}, 
  author={Zibin Liu and Cheng Zhang and Xi Zhao and Yunfei Feng and Bingyu Bai and Dahu Feng and Erhu Feng and Yubin Xia and Haibo Chen},
  year={2025},
  eprint={2512.15784},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2512.15784}, 
}

We gratefully acknowledge the open-source projects like MobileAgent, UI-TARS, and Qwen-VL, etc. We also thank the National Innovation Institute of High-end Smart Appliances for their support of this project.