Tree Memory: An Interpretable, Gradient-Free Learning Stack and Its Measured Laws of Memory, Value, Agency, and Governance

We present Tree Memory, a standalone learning component distilled from an agent-based simulation lineage: any stream of numeric vectors in; concepts, crisis paths, an associative map, and a reaction on every input out. The system learns without gradients, keeps every weight readable as a named pattern, and runs on event-driven episodes. Its development was conducted under a pre-registration discipline: every claim was stated as a falsifiable prediction before data, every campaign carried kill conditions, and every result—positive, negative, and null—was published in an append-only record. Across eleven campaign families we measure a body of laws: memory consolidates by duration and aftermath rather than intensity; value written into a representation at sufficient resolution differentiates its bearers from their siblings under equal exposure; a single shared map can serve multiple agents and arbitrate conflicting wills by differentiation; and a finite economy gives rise to allocation, while governance’s value is shown to live strictly above a measured sensory ceiling. We also publish the boundaries: the conditions under which associative dependency recovery is empty, the refractory limits of causal chains visible to an episode-scale map, and the instrument-physics constraints (co-activation stereoscopy, prune cadence, split-feed fragmentation) that any future campaign must budget against. All campaigns are deterministic, standard-library only, and reproduce bit-for-bit across machines.