Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems

This paper introduces Agentic Context Management (ACM), framing agent memory and context cost as lifecycle and architecture problems rather than only storage and retrieval. It decomposes ACM into five primitives: architecting, ingesting, scoping, anticipating, and compacting and consolidation. The paper argues that naive context accumulation creates quadratic token costs, while crude summarization offers linear cost but can produce an accuracy cliff. It describes a reference implementation, Maximem Synap, reporting 92% on LongMemEval and 93.2% on LoCoMo under the configuration in Section 6.