New research from AI startup Goodfire.ai reports evidence that memorization and reasoning in large language models are carried out in separate neural subcircuits. The team published a preprint in October detailing experiments that separate memory from logical processing in transformer architectures.
The researchers describe that removing memorization pathways caused a 97% drop in the model’s ability to recall training data verbatim, while nearly all of its logical reasoning ability remained intact. This suggests a clean mechanistic split between these functions.
In a layer-level analysis of Allen Institute’s OLMo-7B model, the lower portion of weight components showed higher activation for memorized text, while the top fraction responded more to non-memorized content, illustrating distinct neural channels for memory and reasoning.
An even more surprising result was that arithmetic operations appear to share memorization channels rather than reasoning, with math performance dropping to about 66% after memorization circuits were removed, even when other tasks remained largely unaffected. This helps explain why AI often struggles with math without external tools.
The study notes that “reasoning” in AI covers a spectrum, and the surviving memory-free reasoning tasks include evaluating true/false statements and following conditional rules, but deeper mathematical reasoning remains challenging. The authors caution that results may not generalize across all models or prevent sensitive information from being recoverable, underscoring the early stage of this line of work.
