Aimee Project Releases Native Memory Plugins for Qwen 3.8 and Gemma4 Models

The Aimee project announced the completion of Phase 1 of its roadmap, releasing five vLLM plugins for Qwen 3.8 27B and Gemma4 models (E2B, E4B, 12B, and 26B). These plugins allow models to consume external knowledge stores natively, bypassing traditional context windows and avoiding the context-window penalties associated with text injection. The project asserts that this method is dramatically faster than supplying memory as text in larger workloads and is applicable to transformer, derived transformer, and Mamba architectures. This approach aligns with the project's core belief that reasoning should reside in the model while memory resides in the harness, removing parameter count as a hard limit on retained knowledge. Additionally, the project announced initial success in Phase 2, demonstrating self-learning reasoning capabilities in models like Gemma4 26B, which relies on the Phase 1 infrastructure for safe knowledge updates. Future plans include expanding support to DeepSeek and other models, with source code and experiment data to be released upon publication of their paper. The project explicitly noted that while they explored engrams, they are not the right technology for this specific application, though they remain a complementary idea.