nulllabtests/evo-parliament

This project is a research framework for simulating groups of AI agents that govern themselves. Instead of following fixed rules set by a human designer, the agents in this system propose changes to their own rulebook, debate them, and vote on whether to adopt them, so their constitution evolves over time through the group's own decisions. Each agent also carries a kind of associative memory inspired by how biological brains strengthen connections between things that occur together, called Hebbian memory. This memory quietly influences what an agent proposes and how it votes in future rounds, based on what has worked or failed before. On top of the voting and memory system, the agents also evolve genetically across generations. Agents with better reputations and whose behavior lines up well with the group's shared memory are more likely to pass their traits on to the next generation, similar to natural selection. All of this plays out inside a simplified simulated chemistry environment, where simple abstract molecules interact according to fixed mathematical rules. The point of this environment is to give the group's governance decisions real, measurable consequences, such as how complex, diverse, and energy sustainable the resulting system becomes over time. The authors tested four different voting methods, including simple majority rule, quadratic voting, conviction voting, and liquid democracy, running each one many times. They report that quadratic voting produced the most open ended, increasingly complex outcomes by a wide margin, and that adding the Hebbian memory system meaningfully improved outcomes compared to a version without memory. The project is aimed at researchers studying self governing multi agent systems, constitutional AI, and artificial life, and is built as a reproducible experimental framework rather than a ready to use product. It is written in Python and requires Python 3.10 or newer.