https://arxiv.org/abs/2601.02163
https://arxiv.org/abs/2601.02163
MAGMA: 4 orthogonal graphs (semantic, temporal, causal, entity) with query-adaptive retrieval.
EverMemOS: engram-inspired lifecycle — episodic traces consolidate into semantic structures.
Memory as infrastructure, not feature.
MAGMA: 4 orthogonal graphs (semantic, temporal, causal, entity) with query-adaptive retrieval.
EverMemOS: engram-inspired lifecycle — episodic traces consolidate into semantic structures.
Memory as infrastructure, not feature.
其核心亮点超惊艳:仿人脑设计四层架构(代理层=前额叶、记忆层=大脑皮层等),兼顾记忆存储、高效检索与外部交互;支持公私记忆隔离、多智能体协作,从陪伴到复杂任务全场景适配
现已开源可本地部署,云服务版后续上线,彻底解决AI“健忘”的问题
其核心亮点超惊艳:仿人脑设计四层架构(代理层=前额叶、记忆层=大脑皮层等),兼顾记忆存储、高效检索与外部交互;支持公私记忆隔离、多智能体协作,从陪伴到复杂任务全场景适配
现已开源可本地部署,云服务版后续上线,彻底解决AI“健忘”的问题
EverMind has completed its SOC 2 audit, a milestone in our commitment to protecting customer data.
Thank you to Advantage Partners and Vanta for supporting us throughout the process.
Explore our security practices: trust.evermind.ai/
EverMind has completed its SOC 2 audit, a milestone in our commitment to protecting customer data.
Thank you to Advantage Partners and Vanta for supporting us throughout the process.
Explore our security practices: trust.evermind.ai/
(1) EverMemOS: A Self-Organizing Memory Operating System for Structured Long-Horizon Reasoning
(2) LLM2IR: simple unsupervised contrastive learning makes long-context LLM great retriever
🔍 More at researchtrend.ai/communities/RALM
另外也去对比了下 OpenClaw,它的记忆实现也非常有意思。它不追求全量记忆,而是优先解决“记忆什么时候该被用”。它把记忆拆成了三层:全局层、工作区层、任务层,逐层收敛,只在必要的时候才把上下文拉进来。
另外也去对比了下 OpenClaw,它的记忆实现也非常有意思。它不追求全量记忆,而是优先解决“记忆什么时候该被用”。它把记忆拆成了三层:全局层、工作区层、任务层,逐层收敛,只在必要的时候才把上下文拉进来。