Agentic Context Management: Memory and Cost as Architecture Problems

Production AI agents fail primarily due to poor context management—struggling with accumulating conversation histories, large prompts, and ballooning tool outputs—rather than reasoning deficiencies. The paper proposes Agentic Context Management (ACM) as a lifecycle discipline with five core primitives (architecting, ingesting, scoping, anticipating, and compacting) that achieves linear token cost while preserving fidelity, demonstrated through a reference implementation achieving 92-93% accuracy on benchmarks. The approach addresses practical production needs across organizational scopes and introduces new evaluation dimensions like latency, token efficiency, and context-rot resistance.