AI Engram: In Search of Memory Traces in Artificial Intelligence

Researchers have introduced a geometric framework to identify 'AI engrams' in deep neural networks and proposed a new causal model of theory of mind in artificial intelligence, advancing the understanding of memory and social reasoning in AI.

A team of researchers has made significant progress in understanding how deep neural networks preserve memory traces, introducing a geometric framework to identify 'AI engrams' ^[1]. AI engrams are analogous to biological memory units and enable the surgical manipulation of learned knowledge through linear arithmetic. The framework formalizes neuroscientific criteria of specificity, reactivation, sufficiency, and necessity, and has been shown to be causally valid and scalable in experiments ranging from simple MLPs to LLMs. Separately, a new causal model of theory of mind in conflict for artificial intelligence has been proposed, addressing the question of when to engage in mentalizing ^[2]. The model treats theory of mind as a mechanism activated by situational and agent-level conditions and specifies four exogenous variables and five endogenous mediators. Theory of mind is essential for effective human-machine integration, and existing AI-ToM models have focused on how to mentalize, but not when. The primary outcome of the proposed model is epistemic accuracy, which decouples social reasoning from behavioral policy.