Executable Relational Memory: Graph-Organized Learning, Program Composition, and Causal Revision
We present an executable relational memory architecture in which persistent relational units integrate observations, executable programs, mutable interfaces, and causal dependencies within a shared, graph-organized substrate. Each unit contains a stable three-signal nucleus responsible for activation, while its mutable faces accommodate the procedures through which computational competencies are acquired, tested, composed, executed, and revised. An implemented Genesis organization actively selects informative experiments by identifying disagreements among executable hypotheses, acquires functional relationships from observations, and materializes validated programs as persistent computational units. The architecture maintains independently identifiable positive and negative evidence across contextual projections, enabling contradiction handling, causal revision, and dependency-aware knowledge maintenance. We formalize the operational semantics of this architecture and establish a compositional conservation theorem demonstrating the preservation of observable computational values under admissible contextual reorganization, pure-prefix program factoring, and dependency-certified result reuse. Additional results characterize the effects of finite instruction budgets on operational equivalence and establish information-theoretic bounds for independently retractable evidence. Experimental evaluation confirms the integration of these mechanisms. The original V5 implementation passes all 51 regression tests and acquires 144 executable programs with 1,632 successful checks across four independent worlds. A separate evaluation across 12 newly seeded worlds acquires 144 programs and achieves 1,728 correct checks. In an integrated experiment, five distinct computational tasks coexist within a single memory, producing 100 correct held-out outputs following 25 actively selected observations. A direct mathematical verification suite additionally passes 249,786 targeted comparisons. An experimentally observed Genesis transformation demonstrates that preserving computational output does not necessarily preserve executability under finite instruction budgets. Together, these results establish an implemented and reproducible computational architecture in which relational memory serves simultaneously as a substrate for knowledge representation, executable computation, autonomous competency acquisition, structural reorganization, and causal revision, with mathematically specified conditions for preserving observable behavior.
Authors
- Luan Rosa
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24230716
- Primary Topic
- AI-based Problem Solving and Planning
- Type
- preprint