Maybe one could generate autonomous internal "emotion-like" states with variable timescales to condition the output of an LLM, in order to mix "real time" dynam

Maybe one could generate autonomous internal "emotion-like" states with variable timescales to condition the output of an LLM, in order to mix "real time" dynamics and autoregressive token dynamics?