Delta N-gram Enables Forward-Only Continual Writing into LLM Memory
October 23, 2026
Delta N-gram introduces a forward-only write path into a model's native hashed N-gram table using a 4.1M-parameter writer. This method creates a sparse patch to store evidence without modifying the pretrained table, allowing a 0.6B prototype to retain 16/16 facts after a second write using forward passes only.
HOW THIS AFFECTS YOU
●
builderYou can implement non-destructive, low-latency continual learning by writing to sparse N-gram patches.
●
researcherThis method avoids backpropagation during inference by using a gated target and convex combination rule.