A new paper proposes Agentic Context Management (ACM), reframing agent memory as a lifecycle and architecture problem rather than a simple storage-and-retrieval task. The approach decomposes context handling into five primitives — architecting, ingesting, scoping, anticipating, and compacting/consolidation — to manage conversation histories and tool outputs across long agent sessions. The paper reports that validated compaction achieves linear token-cost growth while preserving accuracy, versus quadratic growth for naive accumulation and accuracy loss for crude summarization, with a reference implementation scoring 92% on LongMemEval and 93.2% on LoCoMo.
