RD-Forget Keeps Every Observation but Controls Which Ones Reach the Answer
arXiv 2609.10263 (2026-09-09) separates what a persistent agent stores from what it uses, because a superseded fact misleads a current-state answer while still being required for a historical query. A retained source archive holds all observations and a query-conditioned memory view governs their influence, with a frozen LLM curator extracting evidence, grouping facts into semantic slots, and preserving relations needed for multi-hop reasoning. Same-slot replacement links suppress superseded values in current-state contexts while intent-aware retrieval makes older evidence eligible again, with a rate-distortion formulation sizing the view to a budget. Ablations show configurations lacking forgetting or query conditioning have the largest deficits.
↳ Follow the thread