Description
Processing temporal information in quantum many-body systems usually follows one of two paradigms. Quantum reservoir computing retains an unmeasured part of the evolving state, allowing past inputs to influence later outputs. Quantum extreme learning machines instead fully measure and reset the system at each step, relying only on instantaneous nonlinear features. We study the tradeoff between these regimes in an anisotropic Heisenberg spin chain with a tunable encoding window that controls how many qubits are overwritten with new data at each step. Using Bayesian optimization to identify high-performing Hamiltonians across five temporal benchmarks, we find that the optimal architecture depends strongly on the temporal depth of the target signal. Short-range tasks are well served by the fully reset limit, whereas tasks with useful structure beyond the encoded window benefit from retaining unmeasured qubits, often with the best performance at intermediate window sizes. An entanglement-based diagnostic further shows that memory-demanding tasks favor weakly scrambling dynamics that preserve temporal correlations, while the fully reset limit favors more strongly entangling dynamics that produce a richer instantaneous feature space. These results suggest that both the encoding window and the Hamiltonian should be matched to the timescale of the target signal.