Researchers discover 'AI engrams' for surgical memory manipulation in neural networks

A team led by Jea Kwon at KAIST and the Institute for Basic Science (IBS) has derived a geometric framework to identify and manipulate memory traces in deep neural networks, which they call "AI engrams." Inspired by neuroscientific criteria of specificity, reactivation, sufficiency, and necessity, the researchers formalize a constrained inverse problem that yields a closed-form estimator to isolate individual memory traces from the globally entangled parameters of a trained network. This biologically-motivated solution corresponds to a natural gradient update on the parameter manifold, allowing memories to be surgically added, composed, or erased through simple linear arithmetic—no iterative optimization or retraining required. The paper, accepted as an Oral at ICML 2026, demonstrates the approach on models ranging from simple MLPs to large language models, confirming both causal validity and scalability.

The practical implications are significant: AI engrams could enable on-the-fly customization of pretrained models—removing outdated or harmful knowledge, inserting new facts, or composing multiple domains without expensive fine-tuning. By linking the distributed storage theory of biological memory with the geometry of deep learning, this work opens a new path for interpretable and editable AI. The open-source code release further accelerates adoption. While still early, the method offers a principled alternative to current editing techniques like ROME or MEMIT, and could reshape how we think about model updates, privacy (e.g., forgetting user data), and continual learning.