Neurochemical memory diversification towards liquid reinforcement learning

The creation of biological memory analogs has long been hindered by the fundamental mechanistic differences between solid-state electronics and biological systems. By manipulating biomolecules in appropriate electrolytic environments, aqueous emulation of memory and its diversification has recently emerged. Here, we present real neurochemical memory diversification based on a nascent organic photoelectrochemical transistor and its implementation toward liquid reinforcement learning (RL). By employing neurochemicals H2S and H2O2 as reward and clearance signals, respectively, the device demonstrates reversible switching between strengthened and suppressed memory, reproducing diversified memory behaviors analogous to those in the human brain. As its diversified memory supports adaptive decision-making and strategy updating, it is further applied to a navigation task for efficient RL. This work provides a strategy for aqueous neurochemical memory diversification and highlights liquid-phase RL for advanced neuromorphic applications.