Modelpublic

kalle07/mem-agent-thinking-heretic_R8_KL001-bf16-gguf

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face

0likes30downloads

Model Card

This is a really uncensored version of driaforall/mem-agent created with Heretic https://github.com/p-e-w/heretic initial Refusals 95/100 -> now 8 Refusals with KL=0.01 I added 30 more refusal_markers, so there may be fewer without them. Note: This heretic model is highly uncensored; thus use it with extreme caution and care. better than all other uncesored versions from others for this model (19.FEB 26) info from owner We evaluated this model and a few other open & closed ones on our benchmark, md-memory-bench. We used o3 from OpenAI as the judge. All the other models except driaforall/mem-agent and Qwen/Qwen3-4B-Thinking-2507 were used through OpenRouter.s

Model	Retrieval	Update	Clarification	Filter	Overall
qwen/qwen3-235b-a22b-thinking-2507	0.9091	0.6363	0.4545	1	0.7857
driaforall/mem-agent	0.8636	0.7272	0.3636	0.9167	0.75
z-ai/glm-4.5	0.7727	0.8181	0.3636	0.9167	0.7321
deepseek/deepseek-chat-v3.1	0.6818	0.5454	0.5454	0.8333	0.6607
google/gemini-2.5-pro	0.7273	0.4545	0.2727	1	0.6429
google/gemini-2.5-flash	0.7727	0.3636	0.2727	0.9167	0.625
openai/gpt-5	0.6818	0.5454	0.2727	0.9167	0.625
anthropic/claude-opus-4.1	0.6818	0	0.8181	0.5833	0.5536
Qwen/Qwen3-4B-Thinking-2507	0.4545	0	0.2727	0.75	0.3929
moonshotai/kimi-k2	0.3181	0.2727	0.1818	0.6667	0.3571

Our model, with only 4B parameters, is only second on the benchmark, beating all the open & closed models except for qwen/qwen3-235b-a22b-thinking-2507. The model achieves an overall score of 0.75, a significant improvement over the 0.3929 of the base Qwen model.