EpiCon: Collective Agent Learning through Co-Evolving Multimodal Memory
EpiCon lets multimodal agents share a persistent experience bank without changing the underlying model weights.
The paper uses two separately trained 2B models: one to refine text and visual memory during attempts, and one to organize and retrieve past lessons. Its evaluations cover eleven benchmarks across four multimodal task domains, with multiple backbones and two harnesses. The authors report macro-average gains of 1.7 to 4.9 points over no-memory baselines, plus a 67% to 74% cut in memory-operation time compared with larger memory models. HF Daily Papers' note
The paper uses two separately trained 2B models: one to refine text and visual memory during attempts, and one to organize and retrieve past lessons. Its evaluations cover eleven benchmarks across four multimodal task domains, with multiple backbones and two harnesses. The authors report macro-average gains of 1.7 to 4.9 points over no-memory baselines, plus a 67% to 74% cut in memory-operation time compared with larger memory models. HF Daily Papers' note
score 5