RefCon: Iterative Refinement and Contrastive Memory Extraction for Context-Evolving Agent
RefCon extracts agent memories via self-refinement and contrast, beating unlabeled baselines on AppWorld and BFCL-V3.
RefCon combines sequential self-refinement with parallel self-contrast so context-evolving agents can extract higher-quality memories without gold labels or retraining. Evaluated on AppWorld and BFCL-V3, it reports relative improvements of 21.6% on ACE and 16.6% on ReMe over no-scaling baselines, while diversity-focused DivCon gains 35.5% on ReasoningBank. It outperforms existing label-free baselines, generalizes across model scales, and surpasses ground-truth memory baselines on software-engineering tasks. Performance keeps improving as more trajectories are used, unlike diversity-only scaling, which saturates earlier.
- RefCon extracts memories without gold labels or retraining.
- Relative gains are 21.6% on ACE and 16.6% on ReMe.
- Diversity variant DivCon gains 35.5% on ReasoningBank.
- It beats ground-truth baselines on software engineering tasks.
- Gains continue as more trajectories are added.
From the original article
1 min read at arxiv.orgLong-horizon agent interactions generate useful but noisy experience, and retraining models to absorb it is expensive. Context-evolving agents therefore need memory extraction methods that improve with more test-time compute without relying on gold labels. We propose RefCon, which combines sequential self-refinement with parallel self-contrast to extract higher-quality memories without gold labels.
By Ubaidillah Ariq Prathama · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.