TinyEngram

Why we included this project

If you're exploring ways to inject domain knowledge into large models without full retraining, TinyEngram offers a concrete alternative to LoRA. It's an open research implementation of DeepSeek's Engram architecture, an N-gram-style memory module with gated retrieval, applied to Qwen language models and extended to Stable Diffusion for concept injection. The repo includes training scripts, pinned dependencies, logs, and detailed experimental reports that back its central claims: Engram-based memory injection can be more parameter-efficient than LoRA while resisting catastrophic forgetting, and it composes cleanly across text and vision. For researchers and engineers working on memory-injection and fine-tuning techniques, this is a reproducible reference rather than a polished product, useful for understanding how a lightweight memory module can teach a model new subjects without retraining the backbone.