Researchers studying how large AI models such as ChatGPT learn and remember information have discovered that their memory and reasoning skills occupy distinct parts of their internal architecture. Their insights could help make AI safer and more trustworthy.
AI models trained on massive datasets rely on at least two major processing features. The first is memory, which allows the system to retrieve and recite information. The second is reasoning, solving new problems by applying generalized principles and learned patterns. But up until now, it wasn’t known if AI’s memory and general intelligence are stored in the same place.
So researchers at the startup Goodfire.ai decided to investigate the internal structure of large language and vision models to understand how they work.
Mapping AI’s brain
First, the team used a mathematical technique called K-FAC (Kronecker-Factored Approximate Curvature) to identify specific processing components responsible for different capabilities, specifically rote memorization in low-curvature pathways (narrow, specialized memory lanes) and flexible reasoning in high-curvature areas (broad, shared processing components).
Then they switched off parts of the AI associated with memorization and tested the model on different tasks. These included answering factual questions and solving new problems. This allowed them to show that when memory was disabled, the models could still use their reasoning skills, indicating that the two functions occupy separate parts of the AI’s internal architecture.
“Our curvature-based pruning approach effectively mitigates memorization the best across both model sizes without requiring supervised training data, achieving notably better generalization to unseen memorized content,” wrote the researchers in their paper published on the arXiv preprint server.
The process of disabling memory revealed a surprising trade-off. While general problem-solving remained intact, the skills AI used for mathematics and recalling isolated facts were heavily affected. “Arithmetic and closed-book fact retrieval rely more on low-curvature directions and are disproportionately impacted by edits, whereas open-book and non-numerical logical reasoning are largely preserved or occasionally improved,” said the authors.
Making AI safer
Knowing exactly how AI works will be key to improving safety and increasing public trust. One problem with AI models that memorize data is that they may leak private information or copyrighted text. Also, this memorization can lead to the retention of harmful biases or toxic content.
But these issues can be mitigated if engineers can precisely target and remove rote-memorized facts and specialized pathways without affecting AI’s general intelligence. Understanding these memory pathways could also make AI models more efficient and less expensive to run by reducing the amount of network space they need.
Written for you by our author Paul Arnold, edited by Gaby Clark, and fact-checked and reviewed by Robert Egan—this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive.
If this reporting matters to you,
please consider a donation (especially monthly).
You’ll get an ad-free account as a thank-you.
More information:
Jack Merullo et al, From Memorization to Reasoning in the Spectrum of Loss Curvature, arXiv (2025). DOI: 10.48550/arxiv.2510.24256
arXiv
© 2025 Science X Network
Citation:
Mapping AI’s brain reveals memory and reasoning are not located in the same place (2025, November 11)
retrieved 11 November 2025
from https://techxplore.com/news/2025-11-ai-brain-reveals-memory.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no
part may be reproduced without the written permission. The content is provided for information purposes only.
