driaforall/mem-agent

driaforall/mem-agent: An Agentic LLM for Memory Management

The driaforall/mem-agent is a 4 billion parameter model built upon Qwen3-4B-Thinking-2507, fine-tuned with GSPO (Zheng et al., 2025) to operate as an agent interacting with a structured, Obsidian-like memory system. Its core strength lies in its ability to manage and utilize information within this memory framework through a set of defined tools.

Key Capabilities

• Memory Interaction: Trained for efficient retrieval, updating, and clarification of information within its memory system.
• Tool Use: Integrates with file and directory operations (create_file, read_file, update_file, delete_file, list_files, etc.) to manipulate its memory.
• Structured Reasoning: Employs <think>, <python>, and <reply> tags for structured agentic loops, executing Python code in a sandbox and processing results.
• Filtered Retrieval: Can handle user-defined <filter> tags to refine or obfuscate retrieved information.
• Benchmark Performance: Achieves an overall score of 0.75 on the md-memory-bench benchmark, outperforming many larger open and closed models, including its base Qwen model, and ranking second only to qwen/qwen3-235b-a22b-thinking-2507.

Good For

• Agentic Applications: Ideal for building agents that require persistent memory and the ability to interact with a file-based knowledge base.
• Knowledge Management Systems: Suitable for systems needing intelligent retrieval, update, and clarification functionalities over structured markdown memory.
• Integration with Larger Models: Recommended for use as an MCP (Memory Control Plane) server, allowing a larger model to delegate memory interaction tasks to mem-agent.