Overview of our approach. Collect activations and gradients from the training data sample (a). This allows us to approximate the loss curvature with respect to the weight matrix using K-FAC (b). Decompose these weight matrices into components, each of the same size as the matrix, ordered from highest to lowest curvature. Language models show that data from different tasks interact with parts of a spectrum of components in different ways (c). Credit: arXiv (2025). DOI: 10.48550/arxiv.2510.24256
Researchers studying how large-scale AI models like ChatGPT learn and remember information have found that their memory and reasoning skills occupy different parts of their internal architecture. Their insights could help make AI more secure and reliable.
AI models trained on large datasets rely on at least two key processing functions. The first is memory, which allows the system to retrieve and recite information. The second is reasoning, which applies generalized principles and learned patterns to solve new problems. But until now, it was unknown whether an AI’s memory and general intelligence are stored in the same place.
So researchers at startup Goodfire.ai decided to look into the inner workings of large-scale language and visual models to understand how they work.
Mapping the AI brain
First, the team used a mathematical technique called Kronecker Factor Approximate Curvature (K-FAC) to identify specific processing components responsible for different functions, specifically memorization in low-curvature paths (narrow specialized memory lanes) and flexible reasoning in high-curvature regions (broad, shared processing components).
They then turned off parts of the AI related to memory and tested the model on a variety of tasks. These include answering factual questions and solving new problems. This shows that the model can use its inference skills even when memory is disabled, and that the two functions occupy separate parts of the AI’s internal architecture.
“Our curvature-based pruning approach effectively maximizes memory reduction across both model sizes and achieves significant generalization to unseen memory content without the need for supervised training data,” the researchers said in a paper published on the arXiv preprint server.
The process of disabling memory revealed some surprising trade-offs. While general problem-solving abilities remain the same, the skills the AI uses to remember math and individual facts have been significantly impacted. “Arithmetic and closed-book fact retrieval rely on low-curvature directions and are disproportionately affected by editing, whereas open-book and non-numerical logical reasoning is largely preserved or, in some cases, improved,” the authors said.
Making AI safer
Knowing exactly how AI works is key to improving safety and increasing public trust. One of the problems with AI models that remember data is that personal information or copyrighted text can be leaked. This memorization can also lead to harmful bias and retention of harmful content.
However, these problems can be alleviated if engineers can precisely target and remove rote facts or specialized pathways without affecting the AI’s general intelligence. Understanding these memory paths can also improve the efficiency of AI models and reduce their execution costs by reducing the amount of network space required.
This article, written for you by author Paul Arnold, edited by Gabby Clark, and fact-checked and reviewed by Robert Egan, is the result of careful human labor. We rely on readers like you to sustain our independent science journalism. If this reporting is important to you, please consider making a donation (especially monthly). As a thank you, we’re giving away an ad-free account.
Further information: Jack Merullo et al, From Memorization to Reasoning in the Spectrum of Loss Curvature, arXiv (2025). DOI: 10.48550/arxiv.2510.24256
Magazine information: arXiv
© 2025 Science X Network
Citation: AI brain mapping reveals memory and reasoning are not in the same place (November 11, 2025) Retrieved November 12, 2025 from https://techxplore.com/news/2025-11-ai-brain-reveals-memory.html
This document is subject to copyright. No part may be reproduced without written permission, except in fair dealing for personal study or research purposes. Content is provided for informational purposes only.
