Thinking
Ideas made inspectable.
Papers, technical articles, research notes, and considered reflections on software, memory, and intelligent systems.
Published thinking
Legacy-First Design: Building Software That Outlives Its Technology Stack
Legacy-First Design treats time as a primary architectural constraint. By separating the permanent identity of a system from its replaceable technologies, LFD proposes a way to build software that remains coherent long after its original stack, team, and context have changed.
A Memory Can Be Stored Without Being Trusted: Inside DREAM v3
DREAM v3 separates memory retention from cognitive influence, introducing provenance, contradiction handling, epistemic hygiene, controlled context construction, and auditable decision traces for persistent AI memory.
Software Contracts Show Structure. Raygon Makes Dependency Explicit.
Traditional software contracts describe what is available, but rarely reveal what each consumer actually depends on. Raygon explores a projection-based model in which dependency, compatibility, and evolution become explicit and observable.
DREAM: Dynamic Retention Episodic Architecture for Memory: Governed Memory through the Separation of Retention and Influence
DREAM v3 presents a model-independent reference architecture for governed episodic memory in artificial intelligence agents. The architecture addresses a central risk of persistent memory systems: information that remains stored may become outdated, contradicted, weakly grounded, or recursively reinforced, yet still be retrieved and treated as authoritative context. The proposal separates memory retention from query-conditioned memory influence. A Retention Policy Engine determines whether an Episodic Unit should remain active, move between storage tiers, be consolidated, archived, or become eligible for pruning. An independent Influence Policy Engine evaluates whether a retrieved memory may affect the current reasoning process and with what level of authority. Retrieval therefore produces a candidate memory rather than automatic authorization. DREAM v3 introduces versioned Episodic Units, auditable meta-memory, provenance and compression lineage, contradiction handling, temporal validity, epistemic recalibration, governed context construction, and immutable decision traces. A state-aware Hygiene Gate acts as the single authorization point between retrieval and context injection, allowing the system to authorize, qualify, arbitrate, reduce, quarantine, consolidate, or block memories according to their epistemic state. The accompanying reference implementation includes an interactive governance playground, automated tests, synthetic benchmarks, and component-ablation experiments. In a controlled synthetic evaluation, the complete architecture achieved 100% grounding fidelity, 0% false reinforcement, 0% continuity loss, and 100% quarantine recall. Removing the separation between retention and influence produced the largest degradation, reaching approximately 34.3% false reinforcement. These results provide architecture-level evidence in a controlled environment. They do not establish generalization to natural conversations, arbitrary language models, multilingual conditions, adversarial workloads, or production-scale deployments. DREAM v3 is presented as an executable and falsifiable research prototype, not as a universal truth-determination system or a production-ready memory service.