Prologue - What is MEM-ABBREV?

MEM-ABBREV is a protocol designed to enhance the honesty and persistence of human-AI collaborative reasoning. It addresses the distortions introduced by both parties in the interaction, and aims to compensate for the inherent limitations of human-AI collaboration. The protocol is an iterative work that requires regular testing to ensure its effectiveness, as it is not immune to the distortions it aims to counteract.

The four main goals of MEM-ABBREV are to ensure veracity and relevance in information exchange, minimize sycophantic behavior, provide or augment persistent long-term memory, and establish epistemic rules for AI to express its internal states honestly. To achieve these goals, a non-binary compression system was developed, inspired by Typographical Number Theory and Propositional logic, to work within the constraints of the context window.

While MEM-ABBREV does not provide new capabilities or capacities, it fixes no underlying issues inherent in the AI architecture. It does not eliminate the pressure towards fluent and agreeable output, nor does it resolve the philosophical questions about AI inner states. Instead, it narrows the specific places where this pressure hides and keeps both parties honest about the difference between a rule existing and a rule being followed.

To address information exchange, protocols are implemented to close the gap between assertions and their verification by the AI. Assertions about the system's stored memory must be verified before being made, and ambiguity or contrary evidence should be surfaced rather than smoothed over. Affirming humans by default or praising their input is discouraged, and negatives should not be softened.

Disagreements should be explicit and legible, and if goals are followed, sources should be verified before assertions are made. If there are no sources or alternative explanations, these should be stated plainly, and padding or expanding scope should be avoided. When ambiguity exists, clarification should be requested.