New Framework Identifies "AI Engrams" for Surgical Memory Manipulation in Neural Networks.

This study proposes a novel geometric framework designed to pinpoint specific memory traces, termed "AI engrams," within deep neural networks. These engrams are conceptualized as artificial counterparts to biological memory units, addressing the question of how learned information is stored and accessed in AI systems. The framework translates neuroscientific criteria for memory into a solvable inverse problem, yielding a closed-form estimator that can isolate individual memory traces from the network's globally distributed parameters. A key implication of this discovery is the ability to surgically manipulate a network's learned knowledge. Researchers can now compose or erase specific subsets of memories using simple linear arithmetic, bypassing the need for complex, iterative retraining processes. This capability has been demonstrated across various models, from simple MLPs to large language models, confirming the causal validity and scalability of the AI engram concept. The findings bridge the gap between theories of biological memory and artificial representation learning, offering deeper geometric insights into how deep networks manage both functional specificity and distributed information storage.