Figure 2. Memory representations occupy regions of state space during experience and learning.

Example trajectories of network states during recollection of two memories, depicted on a neural subspace. The network state is expressed in the firing activity of large neuronal populations. During situations where the environment or context is relatively stationary, the network state exhibits slow drift that constrains learning and plasticity locally (dotted black circles). Upon a major contextual shift, the network state responds with a fast, commensurate shift to a new regime (from Memory 1 to Memory 2, green trajectory) that recalls another memory. The network resides there until another contextual shift kicks it back to Memory 1. In experimental conditions, these contextual shifts are usually experimenter-defined (e.g., placing animals in different enclosures). However, in the wild, they may be shaped by major changes in the animal’s surroundings. In humans, and probably in non-human animals as well, contextual shifts may be internally motivated (i.e., spontaneous recall). Compartmentalization in network state space ensures that learning does not corrupt existing memories while allowing mechanisms for memory modification within local state space regions.