A theoretical study connects two fields previously studied in isolation: diffusion models, the engine behind AI image generators, and associative memory networks, the classic model of how the brain stores and retrieves patterns. The authors show that diffusion models can be understood as memory systems that balance two opposing capabilities: memorising concrete examples and generalising to new data. That balance explains why these architectures perform so well outside the training set.

Associative memory has been a staple of statistical physics for decades. Networks such as Hopfield's store patterns as stable states and retrieve them from incomplete or noisy inputs. The open problem has always been the same: each added pattern degrades the ability to recall previous ones, and memorising too much prevents generalisation. Diffusion models, by contrast, learn to gradually invert a process that adds noise to data until it becomes a Gaussian distribution. The formal link between the two frameworks had not been established.

The team translates the diffusion process into an energy language: each step of the model corresponds to moving across an energy surface whose minima represent the memorised patterns. In this framework, memorisation amounts to creating deep, narrow minima around the training data, while generalisation requires broad minima covering entire regions of data space. The effective temperature of the diffusion process controls which regime dominates: lowering it favours faithful reproduction of examples, raising it favours the generation of new samples.

The main consequence is an explanation of why diffusion models avoid the overfitting typical of other generative networks: the noise injected during training acts as a regulariser, preventing the network from memorising individual examples. The authors argue that this reading allows model behaviour to be predicted from a few parameters, rather than relying solely on empirical experimentation. The proposal paves the way for designing architectures with explicit control over the fidelity-versus-creativity trade-off, and suggests bridges with statistical learning theory and the physics of disordered systems.