Tan et al.
Agent memory narrows LLM response diversity, limiting user access to information
Tan et al. introduce MemBubble, a benchmark of 1000 real conversation instances, showing response entropy drops 34.7% when memory processes preferences.
Chinese Tech
Jiajun Tan · Fei Sun · Xinyu Ma · Shixuan Zhang · Du Su · Qi Cao · +3 more
State Key Laboratory of AI Safety, Institute of Computing Technology, CAS · Baidu Inc. · University of Chinese Academy of Sciences
Research Digest··2 min read
Tan et al.
§