Persistent language agents must retain experience across time, yet each response only needs evidence that matches the current query. A superseded fact can mislead a present‑state answer while remaining essential for historical queries. To address this, we introduce RD‑Forget, a training‑free framework that decouples what is stored from what is used. A retained source archive keeps all observations, and a query‑conditioned memory view decides how those observations influence the current answer. A frozen language‑model curator extracts relevant evidence, groups facts into semantic slots, and preserves the relations required for multi‑hop reasoning. Same‑slot replacement links suppress superseded values in the current context, while intent‑aware retrieval makes earlier evidence eligible again when needed. A rate‑distortion formulation guides the construction of the answer‑time view within a fixed memory budget. Experiments span conversational memory, knowledge updating, fact consolidation, long‑context reasoning, and personalization under a shared answering pipeline. Results show that accurate answers depend on both query‑relevant evidence construction and control over obsolete alternatives. Configurations without forgetting or query conditioning suffer the largest score drops, whereas slot grouping, historical access, and relation preservation provide complementary benefits. Retaining history while selectively controlling its use offers a practical way to handle changing facts and future questions.
Review