What Should an Agent Forget? Separating What Is Stored from What Is Used
RD-Forget keeps an agent’s history intact while limiting which memories can shape a given answer.
The paper describes a training-free framework with a retained source archive and a query-conditioned memory view. Superseded facts are suppressed for current-state questions, but can be retrieved again when the user asks historically. Its curator groups evidence into semantic slots and preserves relations needed for multi-hop reasoning. Experiments report the largest deficits when forgetting or query conditioning is removed. ArXiv · AI/CL/LG's note
The paper describes a training-free framework with a retained source archive and a query-conditioned memory view. Superseded facts are suppressed for current-state questions, but can be retrieved again when the user asks historically. Its curator groups evidence into semantic slots and preserves relations needed for multi-hop reasoning. Experiments report the largest deficits when forgetting or query conditioning is removed. ArXiv · AI/CL/LG's note
score 4