Memory alignment

Memory alignment scores how well a candidate proposal is supported by retrieved memory evidence. The memory agent retrieves relevant memories for each agent role during the deliberation phase; alignment measures how many of those memories point toward the same proposed action or interpretation rather than diverging or contradicting. The cap at five is a deliberate design choice: uncapped alignment would allow a proposal backed by twenty low-quality memories to outperform one backed by three high-quality ones, rewarding retrieval quantity over retrieval precision. Capping at five means the maximum alignment contribution is the same whether a proposal is backed by five memories or fifty - the differentiating factor then comes from the quality and strength of those five, which flows through the reinforcement and trust score signals on the individual memories rather than through raw count. The other three arbitration dimensions - coherence, predicted reward, and risk - are weighted alongside alignment, with the relative weights themselves being part of the policy state that can drift under reinforcement feedback.