HF Today’s top spot is VoiceMem, with 102 votes.

What this work does is to equip a real-time conversational agent with two memory systems: one handles the current dialogue, and the other manages historical preferences, with streaming incremental updates that don’t drop frames.

On-chain agents have been stuck on this exact problem. Over a week, once you’ve talked with which users’ preferences and which portfolio-position preferences, those need to be stored somewhere that can be called across sessions. Pure on-chain is too expensive, while pure cloud is too centralized.

The subnet agents for $NEAR and $TAO face the same challenge. Once the memory module can be shared in a streaming way, agent collaboration isn’t about taking turns feeding context anymore—it’s about directly aligning state.

Engineering progress like this isn’t an immediate win for encryption; it’s more like laying groundwork for the long term. Looking back three years from now, whether on-chain agents can remember things—that’s the watershed moment.

#双脑记忆架构 #链上代理协议 #AI