MindMemOS: Self-Evolving Memory System Enhances AI Agent Adaptation.

Researchers have introduced MindMemOS, a novel memory operating layer designed for AI agents that is both portable and capable of self-evolution. This system addresses the limitation of static memory systems by allowing agents to continuously adapt their memory models, organizational strategies, and procedural knowledge through ongoing interaction and use. MindMemOS structures open-world information using a unified entity property timestructure. Key features include scenario-adaptive memory modeling, higher-order pattern discovery, and autonomous memory refinement via its MindMemEvolve algorithm, which optimizes memory schemas through validation-driven evolutionary search. The system also incorporates "dreaming" to consolidate memories by merging redundancies and resolving conflicts, alongside implicit corrective feedback for human-in-the-loop memory revision. Furthermore, MindSkillEvolve transforms agent execution trajectories into progressively refined and reusable skills. MindMemOS demonstrated strong performance, achieving 94.03% accuracy on LOCOMO and 70.63% on PersonaMem, with MindSkillEvolve improving SpreadsheetBench success by 9.2 percentage points.