ExpeRepair is a novel LLM-based program repair framework that continuously accumulates and reuses historical repair experiences via a dual-memory system. It comprises two core components:

• Program Repair Module: Consists of a Test Agent and a Patch Agent that collaboratively handle three key tasks — test generation, patch generation, and patch validation.
• Memory Module: Captures repair trajectories from the Program Repair Module, extracting concrete demonstrations and summarizing high-level repair strategies. These are stored in episodic memory and semantic memory, respectively. During future repairs, ExpeRepair retrieves relevant demonstrations and insights to enhance its repair strategy for new issues.

git clone <repository-url>
cd ExpeRepair-v1.0
pip install -r requirements.txt

Set your API keys as environment variables:

export PYTHONPATH=$PYTHONPATH:$(pwd)
# OpenAI key
export OPENAI_KEY="your-openai-key"
# Anthropic Key
export CLAUDE_KEY="your-anthropic-key"

Following Agentless, ExpeRepair uses a repository structure format to localize suspicious files. You can either preprocess the repositories yourself, or download the prepared data: swebench_lite_repo_structure.zip.

After downloading, unzip and export its location:

export PROJECT_FILE_LOC={folder which you saved}

Finally, set up the testbed by following the instructions from the auto-code-rover project.

1. Run run_reproduce_initial.sh with all instance IDs.
2. Run run_localization.sh with all instance IDs.
3. For instances where reproduction succeeds (i.e., a valid reproducer_X.py script is generated), run generate_initial.py.
4. For the remaining instances, run generate_w_memory.py, which leverages the accumulated memory and continues updating it.
5. Finally, run validation.py and obtain_predictions.py with all instance IDs.

# reproduce the issue
bash inference/run_reproduce_initial.sh

# localize potential bug locations
bash inference/run_localization.sh

# generate candidate patches
bash inference/run_generate.sh

# validate candidate patches and select the best one
python inference/validation.py

# obtain the prediction results
python inference/obtain_predictions.py

• SWE-bench
• auto-code-rover
• Agentless