https://richardvalentine.dev Research papers on governed memory for AI. en-us Richard@RichardValentine.dev (Ricky Valentine) Thu, 22 Oct 2026 10:20:26 GMT https://richardvalentine.dev/papers/observation-to-operational-intelligence https://richardvalentine.dev/papers/observation-to-operational-intelligence Fri, 03 Jul 2026 00:00:00 GMT AI becomes useful inside a company only after the work itself is made observable, legible, and consequence-aware. This working paper proposes the Observation Sprint Method, a framework for converting real organizational work into structured workflows, technical specifications, and AI-augmented operational intelligence. Building from process mining, task mining, workflow analysis, and organizational digital twin concepts, the paper argues that modern AI systems become more useful when companies first make work observable, legible, and consequence-aware. The framework combines computer process mining, which extracts structured event logs from enterprise systems, with people process mining, which captures the human decisions, exceptions, screen activity, informal workarounds, and tacit knowledge that often sit outside formal software systems. DOI: https://doi.org/10.5281/zenodo.21163447 https://richardvalentine.dev/papers/beyond-flat-memory https://richardvalentine.dev/papers/beyond-flat-memory Tue, 31 Mar 2026 00:00:00 GMT Flat memory is not a performance limitation — it is a governance failure. AI memory needs sovereign temporal graphs and emergent per-user topology. We present the Sovereign Temporal Graph with Emergent User Topology (STG-EUT), a memory governance architecture that combines contamination-aware sovereign governance with emergent per-user knowledge topology. Current AI memory systems store user knowledge in flat, ungoverned structures. This paper identifies the Dual Contamination Problem — a self-reinforcing failure mode in which pre-existing data pollution and AI-generated errors coexist in the same store and amplify each other — and presents an architecture addressing it through sovereign temporal graph governance, emergent per-user topology via unsupervised community detection over human-verified foundational memories, and dual-function identity-aware retrieval where emergent clusters simultaneously narrow retrieval scope and constitute the system's evolving structural model of the user's enduring concerns. DOI: https://doi.org/10.5281/zenodo.19350176 https://richardvalentine.dev/papers/mortal-knowledge-protocol https://richardvalentine.dev/papers/mortal-knowledge-protocol Tue, 03 Mar 2026 00:00:00 GMT Knowledge must be mortal to remain valuable. What is not used should not persist. We present the Mortal Knowledge Protocol (MKP), a governance framework for AI memory systems built on the thesis that knowledge must be mortal to remain valuable. MKP defines a six-state lifecycle governed by attention-weighted decay, where knowledge that is not retrieved gradually loses salience and migrates from active storage to cold archive. DOI: https://doi.org/10.5281/zenodo.18854480 https://richardvalentine.dev/papers/keos-architecture https://richardvalentine.dev/papers/keos-architecture Sun, 01 Mar 2026 00:00:00 GMT If the model is never granted epistemic authority in the first place, neither hallucination nor model collapse can occur. Large language models deployed in enterprise and scientific contexts face two critical reliability challenges: hallucination and model collapse. Current approaches treat these as separate problems requiring post-hoc mitigation. This paper introduces KEOS (Knowledge-Epistemic Operating System), a system architecture that provides immunity to both failure modes by design through a single architectural inversion: the strict reassignment of epistemic authority from the AI model to verified human actors. DOI: https://doi.org/10.5281/zenodo.18828027 https://richardvalentine.dev/papers/crystallization-theory https://richardvalentine.dev/papers/crystallization-theory Mon, 05 Jan 2026 00:00:00 GMT Variance between similar experiences is not noise—it is a signal that reveals universal principles. Existing AI memory systems excel at centralized contexts but struggle when knowledge is distributed across a network of contributors. This paper introduces Crystallization Theory, a conceptual framework for validating knowledge in decentralized networks. We propose three components: 1) Crystallization Theory, a mechanism for knowledge to emerge through confirmation and decay; 2) Communal Inference Reinforcement (CIR), rewarding inference from shared signals over private context; and 3) Core Memory, a two-layer architecture separating agent-discovered patterns from human-provided insight. DOI: https://doi.org/10.5281/zenodo.18150324