mem-agent-f32-gguf/README.md

license, language, base_model, pipeline_tag, library_name, tags

license	language	base_model	pipeline_tag	library_name	tags
apache-2.0	en	driaforall/mem-agent	text-generation	transformers	text-generation-inference agent

mem-agent-f32-gguf

driaforall/mem-agent is an agentic model based on Qwen3-4B-Thinking-2507, fine-tuned using GSPO (Zheng et al., 2025) to interact with an Obsidian-inspired, markdown-based memory system for advanced retrieval, updating, and clarification tasks. It is structured around agentic scaffolding that leverages dedicated tags and tool APIs for file and directory operations, supporting memory filtering and obfuscation, and evaluated on the md-memory-bench where it outperformed most open and closed models except qwen/qwen3-235b-a22b-thinking-2507, with an overall benchmark score of 0.75. The model is designed for use as an MCP server or standalone, and relies on linked markdown files to manage user and entity data, enabling seamless, flexible document-like memory manipulation for agentic or personal assistant scenarios.

Model Files

File Name	Quant Type	File Size
mem-agent.BF16.gguf	BF16	8.05 GB
mem-agent.F16.gguf	F16	8.05 GB
mem-agent.F32.gguf	F32	16.1 GB
mem-agent.Q2_K.gguf	Q2_K	1.67 GB
mem-agent.Q3_K_L.gguf	Q3_K_L	2.24 GB
mem-agent.Q3_K_M.gguf	Q3_K_M	2.08 GB
mem-agent.Q3_K_S.gguf	Q3_K_S	1.89 GB
mem-agent.Q4_K_M.gguf	Q4_K_M	2.5 GB
mem-agent.Q4_K_S.gguf	Q4_K_S	2.38 GB
mem-agent.Q5_K_M.gguf	Q5_K_M	2.89 GB
mem-agent.Q5_K_S.gguf	Q5_K_S	2.82 GB
mem-agent.Q6_K.gguf	Q6_K	3.31 GB
mem-agent.Q8_0.gguf	Q8_0	4.28 GB

Quants Usage

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):