A new paper challenges a key idea in AI memory editing: the belief that removing or adding specific concepts to an artificial neural network works just as well whether you do it all at once or step by step. Ferdinand M. Schessl tested this 'commutative manifold' assumption—which holds that edits commute, like adding numbers—and found it fails under real-world sequential loads.
The original AI Engram method (Kwon et al., 2026) claimed that concept-specific memory traces can be extracted and combined arithmetically, like linear objects. Their Appendix F proposed the Compositional Memory States Hypothesis, suggesting that edited models form a 'commutative manifold' where integrating two concepts yields the same result regardless of order. However, that evidence came only from single or paired edits—single-cycle tests that miss structural fatigue from multiple changes. Schessl ran the authors' own code (TOFU alpha=0.6) across three different AI models from two vendors and two architecture families, with pre-registered predictions. Four key findings replicated in all three models: (1) zero-shot composition and sequential editing diverged by 61–71% of the edit magnitude; (2) edit order is not interchangeable—cutting one Paris landmark before another could determine whether a third, unrelated concept survives; (3) the method's own statistics (layer-input covariances) drifted monotonically with each new edit in every surviving concept; and (4) erased knowledge partially returned under later unrelated edits. These results falsify Appendix F's commutative-manifold hypothesis for sequential editing, while leaving single-edit findings intact. The practical takeaway: for AI unlearning used in compliance, erasure certified today does not guarantee that same erasure holds after the next edit.
Why does this matter for your brain? While this study is about artificial neural networks, the concept of sequential interference echoes how human memory works. Our memories are not static files—they overlap, compete, and can change when we learn or forget new things. This paper shows that even a mathematically clean editing method can't avoid path dependence: the order of operations leaves permanent traces. For your own cognition, it means that the way you learn (sequentially vs. in batches) may affect how well you retain information, and 'erasing' a memory through disuse or replacement isn't clean—traces persist and can resurface.
What you can do: If you're trying to unlearn a bad habit or replace a mistaken belief, don't expect one clean edit—repeated, spaced practice that builds new associations is more reliable. Also, learn related topics together (blocked practice) rather than piecemeal to reduce interference. And if you're curious about the current state of your own cognitive abilities, a free adaptive test can give you a baseline.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
