MemVerse: Multimodal Memory for Lifelong Learning Agents

Researchers have introduced two new frameworks to improve lifelong learning agents: MemVerse, a model-agnostic memory framework, and Skill-enhanced Test-Time Co-Evolution, which enhances an agent's ability to learn from dynamic environments.

MemVerse, introduced in a paper submitted to arXiv on December 3, 2025, and last revised on June 2, 2026^[1], is designed to bridge fast parametric recall with hierarchical retrieval-based memory. This framework supports continual consolidation, adaptive forgetting, and bounded memory growth. It also introduces a periodic distillation mechanism for fast, differentiable recall. Meanwhile, a separate paper submitted to arXiv on June 2, 2026^[2], proposes the Skill-enhanced Test-Time Co-Evolution framework for Online Lifelong Learning Agents. This framework addresses the limitations of existing lifelong learning agents in handling long-horizon tasks by using a two-stage reinforcement learning approach. It includes Verifier-Guided Skill Learning and Online Skill Internalization to enable the agent to learn from its experiences. Experiments on LifelongAgentBench showed that the Skill-enhanced Test-Time Co-Evolution framework improved average performance by 7 absolute points compared to existing baselines^[2].