MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations

MemOps reframes long-term conversational memory as a lifecycle of explicit operations rather than a static fact store or a final-answer task. It introduces structured traces for remembering, forgetting, updating, reflecting, and composed operations, recording triggers, targets, scopes, state transitions, and supporting evidence. A controllable generation pipeline embeds these events into long task-oriented conversations and evaluates systems with six operation-level probe categories under adjacent-evidence and long-context settings. The abstract reports that session-level retrieval outperforms turn-level retrieval, while long-context models struggle to reconstruct ordered memory-state trajectories.