Large language models have been able to talk with some fluidity, have been shown to possess a rudimentary world model, and have some simple abilities to reason given careful prompting. In parallel, self-situated learning techniques have shown that machines can learn to build hypotheses about the world and discover new correlations through experimentation. We propose a system which interconnects LLMs in such a way that learning from experience is enabled. We propose using an open-source LLM (such as Llama 3.1 70b) which allows for incremental model training and have potentially over a dozen instances interacting together. We reference AutoGPT, AgentCoder, Curiosity Driven Red Teaming, and other such systems to illustrate how LLMs can be effectively interconnected. Our proposed system mimics many of the features of Gary Drescher’s schema mechanism based on Piaget’s work: one agent will propose goals to satisfy curiosity-based exploration; multiple agents will propose plans; when the plans differ but one succeeds in reaching a goal, an agent will create a new training sample; one agent may be ‘thinking subconsciously’ to create context other agents may use in a manner not unlike sub-activation, etc. Iterative training will be used to improve the single model as new hypothesis about the world are discovered.