Hugging Face researchers have published a paper arguing that production AI agent failures are often due to poor context management rather than reasoning limitations. They propose a new discipline called Agentic Context Management (ACM), which treats agent memory as a lifecycle and architecture problem, not merely a storage-and-retrieval issue.
The authors decompose ACM into five primitives: architecting, ingesting, scoping, anticipating, and compacting & consolidation. These operate across a hierarchy of user, customer, and client scopes. The paper makes an economic case: naive context accumulation leads to quadratic token cost growth, while crude summarization achieves linear cost but suffers an accuracy cliff. Only validated compaction yields linear cost with preserved fidelity.
The reference implementation, Maximem Synap, is a multi-tenant service that realizes the five primitives. It reports 92% on LongMemEval and 93.2% on LoCoMo using a smaller answer model (gpt-5-mini). The paper also highlights dimensions not captured by existing benchmarks, such as latency, token efficiency, and context-rot resistance.