MemoryAthena studies whether agent memory can be usefully generated rather than only retrieved, combining direct retrieval from an explicit memory table with two generative pathways arbitrated by a lightweight causal routing head trained on counterfactual future-token likelihood advantages. At inference the router either lets a generated candidate modify the retrieval pathway through bounded interpolation, or falls back to direct retrieval exactly. Adding only about 201M memory-side parameters to a frozen backbone, it raises QA accuracy and general-NLP accuracy over the baseline across benchmark averages.
