System of Context: From Memory as Storage to Memory as Formation - A Technical Architecture for Experience-Driven Agent Intelligence

Current AI agent memory systems treat context as a retrieval problem. An agent stores past events and retrieves semantically similar records to inform future decisions. This paper argues that retrieval is necessary but fundamentally insufficient as a foundation for intelligent agents that must operate consistently, improve over time, and earn trust in high-stakes environments. HYPOTHESIS: Agent intelligence is not a function of what an agent can retrieve from memory, but of what it has become through accumulated experience. The correct unit of analysis is not the memory store but the formation pipeline, the computational process by which discrete experiences are transformed into the principles and values that shape future judgment. We propose the System of Context as a new infrastructure category, the third major layer in the history of data management, following RDBMS for structured relational data and NoSQL for unstructured distributed data. Where those categories manage the properties of stored facts, the System of Context manages the properties of accumulated experience: its temporal decay, its semantic compression, its formation into principle, and its expression as organisational identity. The paper presents a five-layer formation architecture - Events, Episodes, Patterns, Principles, and Values, with precise data models and computational mechanisms for each transition. We introduce a dual-origin values model that separates instilled constitutional values from experientially emergent operational values. We identify seven technical problems that current systems have not addressed and propose mechanisms for each. The architecture is validated through domain pressure tests in financial services (SME credit decisioning under macroeconomic environment shift) and healthcare (clinical sepsis risk stratification under population shift and novel pathogen emergence). The domains chosen for their regulatory rigour, outcome complexity, and high cost of agent failure. The central claim is falsifiable and precise: an agent operating with a System of Context will make decisions that are more consistent, more explainable, and better calibrated to its operational environment than an agent operating with retrieval-only memory and that improvement will compound over time as experience accumulates into principle. KEYWORDS: AI agent memory, context formation, experience-driven learning, agent identity, provenance, explainability, dual-origin values, runtime principle formation, system of context, agent infrastructure, financial services AI, clinical decision support