“What if a model could forget without losing its mind?” That question now has a technical foothold, thanks to new research from Goodfire.ai that reveals a clean architectural split between memorization and reasoning in large language models. The work offers a rare mechanistic view into how these systems store and process knowledge and could reshape approaches to AI interpretability, privacy compliance, and performance optimization.
The team examined two different sets of components of the weights that activate strongly for memorized content versus those that dominate during general reasoning in transformer-based architectures such as Allen Institute’s OLMo-7B. Ranking weights by curvature in the model’s loss landscape-that is, a measure of how sensitive performance is to small parameter changes-they found that the bottom 50 percent of components showed 23 percent higher activation on memorized data, while the top 10 percent favored non-memorized text by 26 percent. This ranking was computed using Kronecker-Factored Approximate Curvature (K-FAC), a second-order method which efficiently estimates curvature without prohibitive computational cost.
Removal of the low-curvature, memory-intensive components reduced verbatim recall by 97 percent while preserving 95–106 percent of baseline performance on Boolean logic, if–then reasoning, and commonsense inference benchmarks. The separation was robust across architectures, including 1 billion–parameter OLMo variants and custom 86 million–parameter Vision Transformers trained with intentionally mislabeled ImageNet data. In vision models, excising memorization pathways restored 66.5 percent accuracy on previously mislabeled images, confirming that the phenomenon is not modality-specific.
One of the most striking observations was that arithmetic operations aligned with memorization circuits, not reasoning pathways. For example, after removing memory but keeping the chains of reasoning intact, math task accuracy dropped to 66 percent. Similar results were found in studies involving causal interventions with GPT-J and Pythia, where researchers said: “basic arithmetic ability lives in the memorization pathways, not logic circuits.” This brittleness implies that contemporary LLMs generally draw arithmetic facts from static tables as opposed to performing some sort of compositional computation-a limitation that has direct implications for the integration of tools and neuro-symbolic hybrid designs.
The curvature analysis also showed a spectrum of task dependencies. The closed-book fact retrieval and rare factual queries degraded sharply, such as up to 78 percent, while common facts, like country capitals, were largely unaffected. On the other hand, open-book question-answering relies not on internal recall but on context provided, and this remained entirely robust. This aligns with findings from mechanistic interpretability that show the sharing of reasoning mechanisms across many inputs, with consistently high average curvature, while memorization relates to “idiosyncratic sharp directions” tied to specific examples.
From a governance perspective, the ability to surgically suppress memorization without retraining might be very useful for privacy protection and copyright compliance. Most current machine unlearning methods suppress outputs without removing internal representations and can easily be reactivated by fine-tuning on targeted data. Goodfire does not ensure complete erasure-memories can resurface with extra training-but sidesteps the need for explicit examples of the content to remove, outperforming prior benchmarks like BalancedSubnet by 16.1 percent versus 60 percent residual memorization on unseen historical quotes.
The curvature of the loss landscape grounds this methodology within broader interpretability and optimization research. High-curvature directions match critical parameters of commonly applied mechanisms, similar to the “induction heads” identified in transformer circuit analysis. Low-curvature, memory-coupled components are reminiscent of isolated peaks in the landscape, each anchored to a narrow range of triggers. This geometric framing echoes recent advances in abstract reasoning models, where feature extraction and relation modules can be tuned to amplify relevant patterns while suppressing distractors, enabling reasoning without prior memorization. All these have consequences for model architecture design, too.
Industry experiments with domain-specialized LLMs show that the optimization of memory pathways can reduce inference cache by up to 70 percent; hybrid systems may one day strengthen logic circuits independently of memory. These insights stand to yield models capable of handling sensitive data in a much safer way, coupled with emerging hardware like analog processing-in-memory accelerators. Yet challenges remain.
The entanglement of arithmetic with memorization raises questions about how to restructure numerical reasoning in LLMs, and the distributed nature of neural representations means that “clean” separations may blur at larger scales or in more complex tasks. Still, the Goodfire.ai study marks a significant step toward mapping and controlling the internal geography of neural networks-an advance that moves the field closer to deliberate, component-level governance of AI cognition.
