Memory as a Service (MaaS): Rethinking Contextual Memory as Service-Oriented Modules for Collaborative Agents (2506.22815v1)

Abstract: This position paper aims to rethink the role and design of memory in LLM-based agent systems. We observe that while current memory practices have begun to transcend the limitations of single interactions, they remain conceptually grounded in "bound memory" in terms of design concept-where memory is treated as local state attached to specific context or entities, forming "memory silos" that impede cross-entity collaboration. To overcome this architectural bottleneck, this paper proposes the timely design perspective of "Memory as a Service" (MaaS). MaaS advocates decoupling memory from its conventional role as an interaction byproduct and encapsulating it as a modular service that can be independently callable, dynamically composable, and finely governed. At its core, MaaS leverages the duality of memory-its inherently private nature and its potential for public service-to grant memory controlled, on-demand interoperability across entities. This paper introduces a two-dimensional design space defined by entity structure and service type, illustrating how MaaS aligns with current memory practices while naturally extending them to cross-entity collaborative scenarios. Finally, we outline an open research agenda spanning governance, security, and ethical ecosystems, and call upon the broader research community to explore this shift toward service-oriented memory for collaborative agents operating across entity boundaries.