Position: Compatibility-First Design Is Critical for Progress in Agentic Memory

Peyman Hosseini, Ondrej Bohdal, Mete Ozay, Jijoong Moon, Kyenghun Lee, Hyeonmok Ko, Taha Ceritli

TL;DR: Compatibility-first design—shared interfaces enabling heterogeneous memory systems to interoperate—offers a practical path between fragmented development and premature standardization for agentic AI memory systems.

Abstract: This position paper argues that neither uncoordinated fragmentation nor rigid standardisation is desirable for agentic memory systems in machine learning research. Instead, we advocate for compatibility-first design: shared interfaces and minimal common abstractions that enable heterogeneous memory systems to interoperate without constraining internal architectures. Fragmentation leads to wasted effort, harder-to-maintain code, difficulty comparing results, and increased safety risks, while premature standardisation risks hard-coding decisions in a rapidly evolving space. Compatibility at the interface level captures the benefits of coordination: reduced redundancy plus shared benchmarks and safety practices, without imposing a monolithic standard. We examine credible alternative viewpoints favouring fragmentation or early standardisation, discuss their limitations, and identify concrete steps, including community-driven interface specifications, reference implementations, and interoperable benchmarks to realise the compatibility-first agenda.

Primary Area: Research Priorities, Methodology, and Evaluation

Keywords: Agentic AI, Memory Systems, LLMs, System Design

Submission Number: 625