New agentic memory framework uses 118K tokens per query

Researchers at the National University of Singapore have developed MRAgent, a framework that improves memory management for AI agents in long-horizon reasoning tasks. Abandoning the static 'retrieve-then-reason' approach, the system allows an agent to dynamically build its memory by accumulating evidence step by step. Using a 'Cue-Tag-Content' mechanism, MRAgent significantly reduces token consumption and runtime costs compared to other agentic memory approaches. For example, when a user asks 'How did Nate use the prize money when he won his third video game tournament?', the agent extracts initial cues, prunes irrelevant tags, and retrieves only relevant memories, avoiding context window pollution. This active memory reconstruction outperforms passive retrieval methods on benchmarks like LoCoMo and LongMemEval.