Minimal reservoir computing architectures can store multiple chaotic attractors simultaneously. Feed the system training data from several different chaotic systems, and it learns to reproduce all of them—the Lorenz attractor, the Rössler attractor, whichever strange attractors you provide. The representational capacity is there. But when you try to switch between attractors based on external cues—give the system a signal saying “now produce Lorenz, now produce Rössler”—it fails. The system stores the behaviors but can't select among them on demand.
This is representational capacity without retrieval. The library has all the books, but there's no catalog. The network's state space contains regions corresponding to each attractor, but it lacks sufficient temporal memory to respond to the switching signal. The cue arrives, but the system can't use it to navigate between stored patterns. Switching requires not just encoding multiple behaviors but maintaining enough short-term context to know which one is currently requested.
Note (2026-11-14; read against the paper's abstract and its closing remarks): the essay says some topologies stored well and others switched better, and that none dominated. The abstract says the authors do not observe a robust dependence of multi-attractor performance on reservoir topology and that no single one of the ten topologies consistently outperforms the others for either storage or cue-dependent selection; it says minimal architectures can successfully store multiple chaotic attractors and struggle with task switching. The essay states as fact that the reservoirs lack sufficient temporal memory; the paper says they may lack the robust temporal memory required for cued transitions and that reliable switching likely demands additional memory or control structure.
Testing across 28 system pairs and ten different reservoir topologies found no architecture that consistently solved both problems—storage and cue-dependent selection. Some topologies stored well, others switched better, but none dominated. The tasks make different demands on the substrate. One requires capacity; the other requires addressability.
This separates two things often conflated: having information and being able to act on it. Storage is about state space geometry—does the system have regions corresponding to the target behaviors? Selection is about controllability—can external signals reliably move the system between those regions? A filing cabinet can store many documents, but without labels and dividers, retrieval collapses. Capacity without addressability is a collection, not a system. The books are there, but you can't find them when you need them.