The authors challenge the usual approach to self-evolving memory, which judges each proposed program revision mainly by its aggregate task score.
Agent memory improves when evolution targets capabilities separately
PrisMem preserves capability-specific gains, then integrates complementary memory programs for stronger performance on million-token histories.
Big Tech
Yaoqi Chen · Yuru Feng · Qianxi Zhang · Baotong Lu · Jianan Lu · Zhirui Wang · +5 more
University of Science and Technology of China · Microsoft · University of California, San Diego
Research Digest··3 min read
Chen and colleagues treat an agent’s memory system as executable code that an LLM can iteratively revise using task feedback.
Why this paper
From Microsoft and 2 others
In one line
Guiding agent-memory evolution by individual capability dimensions rather than overall performance uncovers and preserves improvements, yielding gains on million-token histories.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§