Self-Forming Agents: Individuation through Lived Experience — simulation code

Rather than attempting to "endow AI with consciousness," we propose an AI architecture that forms an experience-dependent, intrinsic structure through continuous interaction with the world and a history substrate operating on dual time scales. This approach offers a verifiable hypothesis grounded in engineering principles, avoiding the need to make claims about consciousness itself.Core Design Principles (4 Points)1. History as "substrate" rather than "weights": Current LLMs operate in two stages: fixed weights established during training and inference based solely on context. Our proposal involves a cycle where history (R) is updated at every step; R alters the subsequent inference and is itself updated. This differs slightly from standard RNNs (as a self-referential substrate remains).2. Short-term self and long-term self (dual time scales): Implementation distinguishes between short-timescale (minutes to hours) and long-timescale (persistent history) components. A "personality" gradually emerges within the same model. In the D×MA framework, R (with decay) corresponds to the short-term self, while C (without decay) corresponds to the long-term self.3. Binding with the world: A design where the structure collapses into a single state if the connection is severed (structural death). The self-structure cannot be maintained without continuous interaction with the world. This approach is highly compatible with Embodied AI and continual learning.4. Learning starting from rank-1: New experiences are initially acquired only along the principal direction, gradually expanding into a broader band. This offers the potential for increased robustness against noise.