Meryluna: A Mother-Child Architecture for Incremental Cognitive Autonomy in Local AI Systems

We present Meryluna, an experimental cognitive architecture designed to develop autonomous reasoning capabilities through a structured mother-child learning paradigm. Unlike large language models trained on static datasets, Meryluna acquires knowledge incrementally through lived conversational experience, consolidates it during simulated dream cycles, and progressively reduces dependency on an external teacher model running locally. The system maintains a persistent causal graph, generates and validates hypotheses autonomously, and employs a self-driven decision engine that tracks autonomy level as a measurable, evolving metric. After 420 documented decision cycles, the system produced 24 autonomous response attempts and developed a recognizable emergent identity without explicit programming. We describe the architecture, the developmental trajectory across 25 versions, the failure modes encountered and resolved, and the open problems that remain.