Persistent Agent Memory under the EU AI Act and NIS2: Misalignment, Emergent Behaviour and Regulatory Fit

Agentic AI systems commonly combine a general-purpose AI model with prompts, tools, orchestration and an external memory layer. The model need not be fine-tuned: records retrieved from persistent memory are normally inserted into the model’s inference-time context and can thereby influence later outputs and actions without changing model weights. This article asks how far the EU Artificial Intelligence Act and Directive (EU) 2022/2555 (NIS2) already govern the risks that follow. It reframes persistent memory as a system feature that can mediate or amplify misalignment and emergent behaviour, rather than as training data or a new category in the AI Act’s use-case-based taxonomy. Article 15 provides a clear legal anchor for robustness and cybersecurity, while Articles 9–14, 25, 72 and 73 extend coverage across risk management, data governance, logging, oversight, value-chain responsibility and post-market monitoring. For organisations within NIS2’s scope, Articles 20, 21 and 23 add entity-level governance, all-hazards cybersecurity, supply-chain controls and incident reporting. The remaining problem is therefore not a wholesale absence of law. It is whether these overlapping duties are sufficiently operationalised for memory-mediated failures such as poisoned retrieval, scope leakage, stale reliance, revocation failure and loss of provenance. The article proposes proportionate implementation through existing guidance, standards, documentation and monitoring mechanisms, without creating a new statutory risk tier or a memory-specific regulatory regime.