In the development of advanced AI language models like GPT-5, two core cognitive functions become evident: memorization, which involves reproducing previously encountered text such as famous quotes or literary excerpts, and reasoning, which entails applying abstract principles to solve novel problems. Recent investigations by the AI startup Goodfire.ai have provided groundbreaking evidence that these two capabilities are managed by separate neural circuits within the AI’s architecture.

Their findings reveal a strikingly clear division. According to a preprint study, when researchers selectively disabled the memory-related neural pathways, the models lost nearly 97% of their ability to recall exact training data verbatim, yet their capacity for logical reasoning remained almost fully intact.

For instance, analysis of Layer 22 in the Allen Institute for AI’s OLMo-7B language model demonstrated that the lower half of the model’s weight components exhibited a 23% stronger response to memorized content, whereas the top 10% were 26% more active when processing general, non-memorized inputs. This functional segregation enabled the team to excise memorization without impairing other cognitive functions.

Arithmetic: A Memory-Dependent Skill in AI

Interestingly, the study uncovered that arithmetic operations share neural pathways with memorization rather than with reasoning. When memory circuits were removed, the AI’s math performance plummeted to 66%, while logical reasoning tasks remained largely unaffected. This insight sheds light on why AI language models often struggle with math despite their sophisticated language capabilities.

This phenomenon mirrors human learners who memorize multiplication tables without grasping the underlying concepts, relying on rote recall instead of computation. Current AI models similarly treat simple equations like “2+2=4” as memorized facts rather than outcomes of logical calculation.

It is important to note that AI “reasoning” encompasses a spectrum of abilities distinct from human reasoning. The preserved logical functions include evaluating true/false statements and applying conditional rules, essentially pattern recognition applied to new data. However, more complex mathematical reasoning, such as proofs or innovative problem-solving, remains a challenge for AI, even when pattern matching is unimpaired.

Implications for AI Data Privacy and Content Control

Advancements in isolating and removing memorized information from AI models could pave the way for eliminating copyrighted content, sensitive personal data, or harmful memorized text without compromising the AI’s transformative reasoning abilities. However, researchers caution that complete erasure is difficult due to the distributed nature of data storage within neural networks, marking this as an early but promising step toward responsible AI data management.

Exploring the Neural Terrain: Loss Landscapes and AI Training

To differentiate memorization from reasoning, Goodfire’s team employed the concept of the “loss landscape,” a visualization of how an AI model’s error rate changes as its internal parameters are adjusted. Imagine a machine with millions of dials; the loss landscape maps the error levels across all possible dial settings. During training, AI models use gradient descent to “descend” into valleys of minimal error, refining their outputs.

By examining the curvature of these landscapes, researchers found that memorized facts correspond to sharp peaks and valleys-areas highly sensitive to small parameter changes-while reasoning abilities correspond to smoother, rolling hills that maintain consistent curvature regardless of direction.

Using the Kronecker Factored Approximate Curvature (K-FAC) method, the team demonstrated that memorized information creates sharp spikes in the loss landscape, which average out to flat regions when considered across many examples. In contrast, reasoning-related pathways maintain moderate curvature, reflecting shared mechanisms across diverse inputs.

Validating Across Models and Modalities

The researchers extended their analysis to various AI architectures, including Allen Institute’s OLMo-2 models with 7-billion and 1-billion parameters, and Vision Transformers (ViT-Base) trained on ImageNet with deliberately mislabeled images to induce controlled memorization. They benchmarked their approach against existing memorization removal techniques like BalancedSubnet.

After excising low-curvature components, the models’ recall of memorized content dropped dramatically from nearly 100% to just 3.4%, while logical reasoning tasks retained 95-106% of their original performance. These tasks included Boolean logic evaluations, deduction puzzles involving relational tracking, object tracking through multiple swaps, and benchmarks such as BoolQ for yes/no reasoning, Winogrande for commonsense inference, and OpenBookQA for science questions requiring fact-based reasoning.

Mathematical tasks and closed-book fact retrieval, however, showed a notable decline, with performance falling to 66-86%. Arithmetic was especially vulnerable; even when the AI generated correct reasoning chains, it failed at the calculation step once memory pathways were removed.

Open-book question answering, which relies on external context rather than internal memory, proved most resilient, maintaining near-full accuracy after editing.

Frequency of Information Influences Neural Allocation

The study also revealed that the neural separation between memory and reasoning varies with the rarity of information. Rare facts, such as names of company CEOs, saw a 78% reduction in recall after editing, whereas common knowledge like country capitals remained largely unaffected. This suggests that AI models allocate neural resources differently based on the prevalence of information in their training data.

K-FAC outperformed other memorization removal methods without requiring examples of memorized content during training. It reduced memorization of unseen historical quotes to 16.1%, compared to 60% with BalancedSubnet.

Vision transformers exhibited similar patterns: when trained with intentionally mislabeled images, distinct neural pathways emerged for memorizing incorrect labels versus learning visual patterns. Removing memorization pathways restored 66.5% accuracy on previously mislabeled images.

Challenges and Future Directions in Memory Editing

Despite promising results, the researchers acknowledge limitations. Current unlearning techniques tend to suppress rather than completely erase information, meaning “forgotten” data can sometimes be reactivated through targeted retraining. Additionally, the precise reasons why mathematical abilities degrade so sharply after memory removal remain unclear-whether arithmetic is truly memorized or shares neural circuits with memorization is still under investigation.

Moreover, some detection methods may misclassify complex reasoning as memorization, and the mathematical tools used to analyze loss landscapes can become unreliable at extreme values. Nonetheless, these challenges do not undermine the practical effectiveness of the editing process.