NoteBank Sparse Episodic Memory for Linear Attention
October 10, 2026
NoteBank is a training-free, plug-and-play interface that adds long-context capability to linear attention models. It uses a sparse episodic memory to archive and retrieve text via the backbone's native features without an external embedding model.
HOW THIS AFFECTS YOU
●
builderYou can extend the context window of linear attention models without retraining.
●
researcherThis challenges the idea that recurrent state capacity is the only bottleneck for long context.