The theory of McClelland et al. (1995) holds that mammalian memory is divided between two interacting networks: one, in the hippocampus, learns quickly and records a few recent episodes exactly; the other, in the neocortex, learns slowly and integrates many cases into abstract, durable knowledge (Buckner, 2024, pp. 160–161, 166–167).
Why two. A network that stores what it knows in its weights and learns from error suffers catastrophic interference: learning the new undoes the old. McCloskey & Cohen (1989) showed this by simulating a classic human experiment: one list of paired associates, then another that links the same stimuli to new responses (1989, pp. 124–128; Buckner, 2024, p. 151). If the news of a lion at a certain watering hole is forced into the network with large weight changes, it can damage the network's knowledge of other animals (Kumaran et al., 2016, p. 515). The way out is interleaved learning: the hippocampus stores the episode and replays it many times, mixed with other experiences, so that the cortex can integrate it gradually, presumably during sleep (Buckner, 2024, pp. 166–167; Kumaran et al., 2016, pp. 518–519).
The revision. Kumaran et al. (2016) amend one point: the cortex can learn quickly what fits a prior schema. Rats familiar with an environment learned two new flavor–place associations from a single exposure (2016, p. 525); slowness is needed only for what contradicts existing knowledge (2016, p. 527).
In AI. The same authors see deep networks as a "neocortical" system of slow, interleaved learning, and the experience-replay buffer of the DQN reinforcement learning agent as a hippocampus: without it, performance fell to at best 30% (2016, pp. 527–528). Buckner draws a further consequence: if humans rehearse each episode hundreds or thousands of times while consolidating it, this should count when comparing how much data they and machines need (2024, p. 167).
For discussion. It is tempting to map the theory onto language models (slow weights, fast context) or onto hybrids, with their compressed and verbatim memories. But in both cases the weights are frozen during generation and nothing passes from one memory to the other: what is missing is precisely consolidation, the heart of the theory.