AI Engram: In Search of Memory Traces in Artificial Intelligence

Memory formation is fundamental to intelligence, yet whether deep neural networks preserve identifiable memory traces analogous to biological memory units remains an open question. This work introduces a geometric framework to identify such “AI engrams” by formalizing the neuroscientific criteria of specificity, reactivation, sufficiency, and necessity into a constrained inverse problem. We derive a closed-form estimator that isolates individual memory traces from globally entangled parameters, and show that this biologically-derived solution corresponds to a natural gradient update on the parameter manifold. AI engrams enable surgical manipulation of learned knowledge: any subset of memories can be composed or erased through linear arithmetic, without iterative optimization. Experiments ranging from simple MLPs to LLMs demonstrate the causal validity and substantial scalability of AI engrams. Together, these results bridge theories of biological memory and artificial representation learning and offer geometric insight into how deep networks simultaneously support functional specificity within distributed storage.

For a century, neuroscientists have searched for the "engram," the physical trace that a single memory leaves in the brain. Artificial neural networks pose the same puzzle: knowledge is smeared across millions of tangled connections, with no map showing where any one fact lives. This makes it hard to delete a specific piece of knowledge, such as private data a model should forget, without retraining the whole system or accidentally wiping out unrelated skills. We turned the criteria that neuroscientists use to identify a memory into precise mathematical rules. Applying them lets us pinpoint the trace of a chosen memory directly from a trained network's existing connections, in a single calculation rather than slow trial-and-error retraining. Because each trace acts like an independent building block, memories can be combined or removed through simple addition and subtraction. We show this works across systems ranging from small networks to billion-parameter language models, cleanly erasing a target concept while leaving the rest intact. The result is a faster and more transparent way to edit what an AI knows, and it reveals a surprising bridge between how brains and machines store memory.