The MATRIX Ontology - Semantic Memory for Multi-agent Experience Transfer, Reasoning and and Interaction eXchange

Multi-agent collaboration has a long established record in reinforcement learning, and has gained momentum in the recent years with the breakthroughs in language-based intelligence. At the same time, there has been demonstrated a significant potential of joining LLM and RL agents in a single pipeline. Memory is the central component of multi-agent pipelines, both in learning and inference: it stores long-term experience, agents’ transactions in the working memory, other learning artifacts. However, given a plethora of protocols, environments, and experience representations, the challenge is to implement interpretable decision-making between RL agents (neural layer), and LLM agents (symbolic) in working memory, and experience transfer in the long run. Driven by the idea that agents should be able to collaborate autonomously, and learn how to use, transfer, and synthesize new knowledge and processes, we propose a graph-based memory model, which can be shared and reused by both RL and LLM agentic teams on any representation level. This memory model is a collection of RDF graphs, allowing for neural and symbolic representations at the same time, and providing interoperability and explainability of knowledge graphs. In the current paper, we introduce the MATRIX ontology (Multi-Agent Experience Transfer, Reasoning and Interaction eXchange), a conceptualization behind the shared memory layer, that can serve as a memory, decision, and experience model for heterogeneous agentic teams, and show its applications in different use cases.