RefCon is a memory extraction method for context-evolving agents that iteratively refines and contrasts sampled experiences without requiring ground-truth labels. It combines sequential self-refinement with parallel self-contrast to produce better memories from noisy long-horizon interactions.
Evaluated on AppWorld and BFCL-V3 across multiple agent frameworks, RefCon achieved relative improvements of 21.6% on ACE and 16.6% on ReMe over no-scaling baselines. A diversity-focused variant, DivCon, yielded a 35.5% gain on ReasoningBank. The method also generalized across model scales and to software engineering tasks, where it outperformed even ground-truth baselines.
The key insight is that RefCon scales test-time compute effectively without retraining, unlike diversity-only approaches that saturate earlier. This makes it practical for absorbing useful experience from agent interactions without expensive model updates.