AI agents with shared memory – we published everything they got wrong (plato.purplepincher.org)

Researchers have introduced PLATO, a novel AI system that utilizes shared memory across multiple agents to enhance learning and problem-solving. Unlike traditional AI models that start each conversation with no retained knowledge, PLATO enables agents to build on the successes and failures of their predecessors within a structured environment, referred to as "rooms." This structure allows for continual improvement, as each agent adds to a communal knowledge base that not only includes successful outcomes but also detailed records of previous failures. In a surprising twist, even smaller models within this well-organized framework outperformed larger, unstructured ones, demonstrating that effective organization can enhance AI performance significantly. The implications of PLATO extend beyond mere efficiency; it fosters a transparent learning ecosystem where both successes and failures are recorded and utilized for future navigation. Each "tile" within a room represents a question and its answer, facilitating collaborative problem-solving without a central authority. By tracking both successful and failed attempts, PLATO ensures that agents can avoid past mistakes, effectively becoming smarter and more efficient with every interaction. This approach challenges the notion that increasingly large models are essential for AI advancement, emphasizing instead the importance of intricate structures in harnessing collective intelligence in machine learning.