{"data":{"slug":"autoark-tinyengram","name":"TinyEngram","tagline":"Research of DeepSeek Engram Architecture based on Qwen-3 and Stable Diffusion series","github_url":"https://github.com/AutoArk/TinyEngram","owner":"AutoArk","repo":"TinyEngram","owner_avatar_url":"https://avatars.githubusercontent.com/u/224494850?v=4","primary_language":"Python","stars":1376,"forks":90,"topics":["deepseek","deepseek-ai","engram","fine-tuning","llm","llm-memory","lora","memory-injection","peft","pytorch","qwen","stable-diffusion","tinyengram","transformer"],"archived":false,"github_pushed_at":"2026-05-21T06:01:13+00:00","maintenance_label":"Slowing","stars_delta_30d":641,"url":"https://www.graphcanon.com/tools/autoark-tinyengram","markdown_url":"https://www.graphcanon.com/tools/autoark-tinyengram.md","api_url":"https://www.graphcanon.com/api/graphcanon/tools/autoark-tinyengram","graph_url":"https://www.graphcanon.com/api/graphcanon/graph?tool=autoark-tinyengram","description":"Research of DeepSeek Engram Architecture based on Qwen-3 and Stable Diffusion series.","homepage_url":null,"license":null,"open_issues":10,"watchers":98,"ai_summary":"TinyEngram repository focuses on research related to the DeepSeek Engram architecture using components like Qwen-3 and Stable Diffusion for tasks such as fine-tuning, model memory injection, and other LLM-related operations.","readme_excerpt":"
\n \"TinyEngram\n\n# TinyEngram: Exploring New Axis of Scaling and Memory Injection\n\n

\n Open research on DeepSeek-AI's Engram and memory injection in Qwen, StableDiffusion and more.\n

\n\n

\n \n \"Qwen\"\n \n \n \n \"DeepSeek\n \n\n \n \"Stable\n \n\n \n \"License\"\n \n\n \n \"arXiv:2605.20309\"\n \n

\n\n[**Report Issues**](https://github.com/AutoArk/TinyEngram/issues)\n\n
\n\n> [!NOTE]\n> **TL;DR:** TinyEngram demonstrates that Engram-based memory injection outperforms LoRA in both **parameter efficiency** and **catastrophic forgetting resistance**β€”and extends seamlessly to **vision** (e.g., Stable Diffusion) for lightweight, composable concept injection. All code, logs, and experiments are open!\n>\n> *If you find TinyEngram useful, a ⭐ helps support the project.*\n\n
\n\n
\n
\n\n πŸ“’ Latest Announcements\n \n\n
\n\n- 2026.05.20 β€” πŸ“ **TinyEngram-Vision technical report is ready.** We organized the vision findings into a complete technical report to invite discussion and further exploration. [Read the report](./doc/paper/tinyengram_vision_paper.pdf), or check out the [arXiv version](https://arxiv.org/abs/2605.20309) if you prefer.\n- 2026.02.12 β€” πŸ–ΌοΈ **TinyEngram meets Vision!** We injected visual concepts into Stable Diffusion through Engram, check our new cross-modal experiment!\n- 2026.02.02 β€” πŸ“Œ Released reproduction scripts for **Engram vs LoRA** experiment.\n- 2026.01.30 β€” πŸ“Œ Added comparison of **catastrophic forgetting** between **TinyEngram and LoRA**.\n- 2026.01.30 β€” πŸ“Œ Added **parameter ablation studies** of TinyEngram with convergence observations.\n- 2026.01.23 β€” πŸŽ‰ Initial TinyEngram commit.\n\n
\n
\n
\n\n
\n\n
\n

\n πŸ” Quick Navigation\n

\n

\n Key Finding 1: Engram as Parameter Efficient Fine-Tuning Method
\n