Neuro-Symbolic AI Boosts Temporal Knowledge-Graph Memory

In partially observable reinforcement learning, agents must intelligently manage their memory over time, deciding which information to retain, retrieve, and discard. This research proposes a novel neuro-symbolic meta-policy designed for temporal knowledge-graph memory. This meta-policy learns to select the most appropriate symbolic memory heuristic at each decision point, while ensuring that the execution remains entirely symbolic. The framework is specifically applied to RoomKG, where hidden states and observations are represented as Resource Description Framework (RDF) graphs, and memory is enhanced with temporal RDF triple annotations. The model integrates knowledge-graph encoding of memory contents with value heads dedicated to question answering, exploration, and forgetting mechanisms. This combination results in a controller that is both adaptive and fully inspectable, providing direct grounding in Semantic Web principles through its RDF-based representation and graph-based symbolic operations. Experiments on train/test room splits, with a long-term memory capacity of 512, show that a qualifier-aware StarE-GNN configuration achieves the best performance among compared symbolic, neural, and neuro-symbolic systems. Crucially, it preserves step-level traceability of all memory-management decisions, offering transparency in complex AI systems.