MemHarness: Memory Is Reconstructed, Not Replayed
Paper • 2607.28272 • Published • 13
MemHarness: Memory Is Reconstructed, Not Replayed
This repository contains the model described in MemHarness: Memory Is Reconstructed, Not Replayed.
Paper: arXiv | Hugging Face Paper
Code: https://github.com/KnowledgeXLab/MemHarness
Description
MemHarness is a framework that equips LLM agents to actively harness and reconstruct past experiences based on the present context — instead of replaying retrieved memories verbatim. This model is a Qwen2.5-7B-Instruct based model fine-tuned with GRPO for memory-augmented decision making in agentic tasks such as ALFWorld and WebShop. It demonstrates state-of-the-art performance in both in-distribution and out-of-distribution scenarios.
Please refer to the paper for full details.
Install from pip and serve model
# Install vLLM from pip: pip install vllm# Start the vLLM server: vllm serve "KnowledgeXLab/MemHarness"# Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "KnowledgeXLab/MemHarness", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'