The NOWA-Cognitive architecture simulates an agent that learns sequentially without backpropagation or a global loss function. Instead of traditional training, the system mimics biological processes where an agent records experiences during the day and consolidates memory during a sleep cycle. The project, developed by Mikhail Burlakov, consists of approximately 125,000 lines of Python code.
The agent, identified as G3A1SF, perceives its environment through context tokens. Its emotional state is represented by a continuous vector across eight axes, including stress, curiosity, and valence. These states are influenced by external events, memories, and internal baselines. Decisions are made through a competition between 'Go' and 'NoGo' channels in the basal ganglia, modulated by neurotransmitters like dopamine and cortisol.
During the day, the agent performs online updates to its value layer based on temporal difference errors, regulated by acetylcholine. However, action weights and long-term memory patterns remain unchanged until the sleep phase. The sleep cycle, managed by SleepManager, filters experiences based on prediction errors and stability, prioritizing events that were surprising or unresolved.
At night, the system performs consolidation by replaying selected episodes. It adjusts action weights, merges similar experiences into cortical patterns, and prunes inactive or rarely reinforced memories. The agent's specific behavioral traits, such as stress decay rates and circadian rhythms, are determined by a genome vector containing 39 genes that define parameters for various cognitive modules.
