Memory Architecture for Managed Cloud Agents: A Memory-Plane Approach to Hydration, Persistence, and Forgetting

Abstract

Memory" has become a catch-all term in applied agent engineering: it covers saved user facts, retained conversation history, workspace knowledge, transient session state, and the retrieval indices used to search all of it. Production systems routinely collapse this variety into a single shorthand, treating memory as retrieval-augmented generation over a vector database. This article argues that the shorthand is an architectural error and proposes an alternative account of agent memory as a memory plane: a brokered, policy-aware subsystem that hydrates prompt context on the read path and performs selective, governed persistence on the write path. Drawing on the memory subsystems exposed by managed-agent products, open-source agent frameworks, and cloud data-management primitives, the article separates a brokered control-and-data plane from backend stores, defines a five-store logical taxonomy together with an authority rule that treats the vector index as a derived structure rather than a source of truth, and specifies a broker interface built around a tombstone-cascade-verify protocol for verifiable forgetting. It frames consistency and durability as per-memory-class decisions mapped to managed services across three major cloud platforms, and it proposes an operational treatment covering engineering service-level objectives, and a low-risk migration path. The latency and service-level figures presented throughout should be read as engineering design targets rather than as measurements or vendor guarantees. The article compares the memory plane against six competing architectures and closes with a set of problems that appear to remain open, among them autonomous consolidation quality, contradiction repair at scale, and cross-agent shared memory.