Geometric Framework Identifies and Erases Individual Memory Traces in LLMs

A closed-form estimator derived from neuroscientific engram criteria isolates individual memory traces in neural networks as constrained inverse problems on the parameter manifold, corresponding to natural gradient updates. The method enables linear-arithmetic composition or erasure of specific memories without iterative optimization, demonstrated from MLPs up to LLMs.