Capability-Driven Self-Evolution of Agent Memory
PrisMem steers memory updates by separate capability targets, then merges the gains into one executable memory program.
The paper argues that holistic feedback can hide useful improvements when gains in one area are offset by regressions in another. Its method selects capability targets with likely spillover effects, diagnoses failures from prior history, and compares candidate programs on paired differential cases. In experiments, PrisMem beats the strongest baselines by 10.54 points on BEAM-1M and 7.83 points on LongMemEval-M, including million-token histories. Source: HF Daily Papers' note.
The paper argues that holistic feedback can hide useful improvements when gains in one area are offset by regressions in another. Its method selects capability targets with likely spillover effects, diagnoses failures from prior history, and compares candidate programs on paired differential cases. In experiments, PrisMem beats the strongest baselines by 10.54 points on BEAM-1M and 7.83 points on LongMemEval-M, including million-token histories. Source: HF Daily Papers' note.
score 4