Embodied AI and wearable assistants require agents to retrieve sparse evidence from continuous days-level videos. Existing memory-based methods typically follow a passive recording paradigm based on fixed-window captioning and periodic memory construction, which produces redundant memories and inefficient long-horizon reasoning. Inspired by human memory formation, we propose Active Explore Memory Agent (AEMAgent), which transforms passive recording into active memory formation by deciding when to remember, how to explore events, and when to reflect. These decisions are instantiated as a coherent pipeline: Surprise-guided Event Selection detects salient event changes for selective memory writing, Exploratory Episodic Memory Encoding enriches accepted events through self-questioning, Predictive-Coding-Guided Reflection updates knowledge from expectation-observation mismatches, and Semantic Memory Consolidation distills noisy episodic updates into compact long-term memory. Experiments on multiple benchmarks show that AEMAgent achieves the best overall performance among all compared methods while maintaining lower storage overhead than strong full-memory baselines, demonstrating the effectiveness of selective and efficient memory formation for scalable very-long video understanding.

Omni-modal Large Language Models (OLLMs) excel in diverse tasks but struggle with complex emotional reasoning, which requires integrating textual, visual, and acoustic signals. We attribute this limitation to modality collapse, where models over-rely on a dominant modality while neglecting complementary cues. To address this issue, we introduce OmniCoT, a data paradigm that interleaves guided tokens (e.g., [vision], [audio]) into reasoning traces to enforce structured evidence extraction. To further internalize the reasoning behaviors instilled by OmniCoT and facilitate adaptive modality prioritization, we propose Dynamic Modality-Entropy GRPO (DyME-GRPO), which utilizes entropy-based uncertainty estimates over Guided Tokens (GTs) to regulate modality usage, thereby mitigating collapse and informational redundancy. By applying supervised fine-tuning with OmniCoT followed by DyME-GRPO, we develop EmoOmni based on the Qwen2.5-Omni-7B backbone. Extensive experiments demonstrate that EmoOmni achieves state-of-the-art performance on multiple emotion recognition and reasoning benchmarks while preserving the general capabilities of the base model. These findings highlight the potential of our work for omni-modal reasoning across a broader range of complex tasks.