Memory Architecture in Compliance-Critical AI Agent System: A survey

Large language model (LLM) agents are increasingly deployed in high-stakes domains—financial crime detection, healthcare records management, and cross-jurisdictional legal reasoning—where regulatory compliance is not optional but legally mandated. In these settings, agent memory is not merely a capability concern but a legal artifact: agents must recall regulated data with precision, enforce entity-level access boundaries, support post-hoc auditability, and selectively erase personal information on demand. Yet existing surveys of LLM agent memory address none of these requirements, and compliance-focused AI papers rarely engage with the architectural vocabulary of modern memory systems. This survey closes that gap by synthesizing twelve recent works—eight compliance-focused and four foundational—into a unified analytical framework organized around the Cognitive Architectures for Language Agents (CoALA) taxonomy of working, episodic, semantic, and procedural memory. We examine how each memory type is adapted—or fails to adapt—to the demands of the Anti-Money Laundering/Bank Secrecy Act (AML/BSA), the Health Insurance Portability and Accountability Act (HIPAA), and the General Data Protection Regulation (GDPR). Our analysis identifies five core architectural tensions (capability vs. auditability, persistence vs. erasure, flat retrieval vs. relational reasoning, session-scoped vs. entity-scoped memory, and semantic drift vs. staleness) and five concrete research gaps: the absence of compliance benchmarks, underspecified audit trail generation, unsolved knowledge graph maintenance, uncharted multi-agent memory propagation, and unquantified true-positive suppression in screening systems. To our knowledge, this is the first survey to synthesize foundational memory architecture research with compliance-domain requirements into a unified analytical framework, and the first to propose a targeted research agenda for the resulting open problems.