Long-term memory architectures for large language models: An enterprise and access control review

Long-term memory architectures for large language models: An enterprise and access control review

Sheyda Sattari Ghahfarokhy1 MohammadReza Bahrani2 Maryam Mazarei3

1) Artificial Intelligence Labratory at NEVISA Engineering Team, Shiraz, Iran Email:
2) Artificial Intelligence Labratory at NEVISA Engineering Team, Shiraz, Iran Email:
3) Artificial Intelligence Labratory at NEVISA Engineering Team, Shiraz, Iran Email:

IOI Code : AAHD-ACZCZK
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    Large language models are inherently limited by their finite context windows, leaving them with no way to retain memory across sessions. This constraint has triggered a massive wave of custom memory architectures, shaped around dialogue scenarios, while data governance, multi-level access control, definitive deletion, and provenance tracking have been addressed in a separate literature. Using a uniform extraction framework, a comparative table and a matrix scoring each system against six criteria of enterprise suitability, this review examines how these architectures are classified, how their lifecycles are implemented, how they are evaluated, and which suit enterprise deployment. In most cases we reviewed, the conversational turn seems to be the dominant unit of scale, rather than the document; the intersection of agentic memory and access governance emerged only in 2025 and remains limited in scale and deletion guarantees. Deletion carries several meanings, only a small minority of systems support physical deletion, and certified unlearning remains unavailable. Eight research gaps, a reference architecture, and five metrics are proposed.

Keywords : retrieval augmented generation agent memory data governance machine unlearning