The authors train a shared policy to manage memory through sequences of tool calls, potentially involving hundreds of operations before an answer receives a score.
Jointly learning what to store and retrieve improves agent memory
Uni-Mem trains one memory-management policy with alternating reinforcement-learning updates and event-level feedback, outperforming the strongest RL-based comparator reported on two benchmarks.
Chinese Tech
Zhiqian Xia · Jielong Tang · Ling-An Zeng · Shixiong Kai · Rong Yao · Hao Kong · +5 more
Southeast University · Huawei Technologies, Noah’s Ark Lab · Sun Yat-sen University · Shandong Normal University
Research Digest··2 min read
Xia and colleagues present Uni-Mem, a system that learns both how to build an AI agent’s long-term memory and how to retrieve information from it.
§