Agentic Context Management — Précis

A Maximem paper arguing that production AI agents fail less from weak reasoning than from unmanaged context: they drown in accumulating history, miss recalls within and across conversations, and pay a token cost that grows with every turn. The incumbent response treats this as a storage and retrieval problem ("memory"). The paper argues that framing is too narrow — actively managing what an agent holds in mind is a lifecycle, not a store.

The author names the discipline Agentic Context Management (ACM) and decomposes it into five primitives: architecting (designing the per-agent memory schema before anything is stored), ingesting (turning raw signals into structured, retrievable memory), scoping (which fraction of knowledge is relevant, at what organizational level), anticipating (prefetching what the next turn will need), and compacting & consolidation (reducing context to fit a budget without dropping what matters). A system that does one primitive well is a memory tool; one that does all five coherently, across scopes, is a context-management platform.

Two arguments make the case quantitative. First, the economic one: full-append context grows token cost quadratically with conversation length, while crude summarization buys linear cost at the price of an "accuracy cliff" (one prior study compressed 18,282 tokens to 122 and dropped accuracy from 66.7% to 57.1%, below the no-context baseline). Only validated compaction — compaction that checks itself for information loss — reaches linear cost with preserved fidelity. Second, the retrieval one: an original five-domain study shows keyword and vector search win in different regimes, and that hit-based scoring cannot detect the "reasoning-sufficiency gap" — retrieving a relevant document while missing the bridge document needed to finish the chain.

The reference implementation is Maximem Synap, a multi-tenant service realizing all five primitives across a scope hierarchy (user → customer → client). It reports 92.0% on LongMemEval and 93.2% on LoCoMo categories 1–4, using a smaller answer model (gpt-5-mini) than competitors — which the author reads as evidence the gain comes from the context layer, not the model. The paper closes by pointing at the frontier: decision-level context (capturing why an organization decided what it did) and the benchmark gaps no current eval measures — latency, token efficiency, and context-rot resistance.