LLM agents require long‑term memory that can retrieve precise and mutually consistent evidence from ever‑growing interaction histories, especially when answers depend on relations among facts rather than isolated semantic matches. Existing memory pipelines remain fragile at two points: indexing errors can fragment repeated entities or merge distinct ones, and online retrieval often treats initial similarity scores as final even when the surrounding graph offers evidence for revision.\
ActionRAG introduces a graph‑memory architecture that couples cautious indexing with query‑time graph retro‑evaluation. The indexing pipeline first reconciles type labels and applies tiered entity resolution to eliminate conflicts. At query time, a dual‑channel retrieval returns both entities and triples, forming an initial evidence set and inducing a bounded local graph. Query‑conditioned receiver‑normalized diffusion then refines triple scores using the local path structure. To preserve direct evidence while recovering missed connections, the system employs seed protection, hub‑penalized flux, and triple materialization.\
Evaluation on MemoryAgentBench (MAB)—which tests accurate retrieval, conflict resolution, long‑range understanding, and test‑time learning across eight deterministic scoring columns—shows ActionRAG achieving an average accuracy of 55.1%, a 5.4‑point gain over the strongest retrieval‑based baseline. Under MAB’s query‑time cost accounting, the mean synthesis cost stays below $0.001 per query on the four reported cost columns, and the sum of separately measured retrieval‑side p99 components is 440.7 ms on a consumer Apple Silicon setup.\
Review: ActionRAG delivers strong performance in long‑term memory retrieval, conflict resolution, and cost efficiency, making it a practical and high‑throughput solution for real‑world LLM agents.