ActionRAG preprint reports 5.4-point gain in agent memory benchmark

A preprint reports that ActionRAG, a system for retrieving evidence from AI agents' interaction histories, achieved 55.1 average accuracy across eight deterministic-scoring columns in MemoryAgentBench. That exceeded the strongest retrieval-based memory baseline by 5.4 points. ActionRAG reconciles entity identities when storing information, then uses relationships among retrieved facts to revise their relevance to a query. Mean synthesis cost stayed below $0.001 per query on four reported cost columns under the benchmark's query-time accounting. Separately measured retrieval-side p99 components summed to 440.7 ms on the authors' consumer Apple Silicon setup; that figure is not an end-to-end latency measurement.