MindMemOS: A Portable and Self-Evolving Memory Operating Layer for AI Agents
06:00 · August 15, 2026 · arXiv cs.AI RSS
Memory is a core component of AI agents, enabling them to accumulate experience, maintain personalization, and adapt over long-term interactions. However, existing memory systems often remain fixed after development, limiting their ability to adapt their memory models, organization strategies, and procedural knowledge through continued use. We present MindMemOS, a portable and self-evolving memory operating layer that organizes open-world information using a unified entity property timestructure. MindMemOS supports scenario-adaptive memory modeling, higher-order pattern discovery, autonomous memory refinement, and continuous skill evolution. Its MindMemEvolve algorithm employs validation-driven evolutionary search to optimize memory schemas for target scenarios, whiledreaming consolidates accumulated memories by merging redundant records and resolving conflicts. In addition, implicit corrective feedback serves as a human-in-the-loop signal for identifying and revising potentially inaccurate or misaligned memories. Its MindSkillEvolve algorithm further transforms agent execution trajectories into reusable and progressively refined skills. MindMemOS achieves 94.03% accuracy on LOCOMO and 70.63% on PersonaMem. MindSkillEvolve improves SpreadsheetBench success by 9.2 percentage points over the initial-skill baseline.
Summary
MindMemOS provides a portable memory layer for LLM-based agents that externalizes long-term information outside any single context window. It organizes open-world data through a unified entity–property–time structure, where each record links a subject entity to a specific attribute or pattern and a temporal reference. This three-dimensional graph supports both unstructured ingestion and scenario-specific schemas while enabling hybrid retrieval across semantic, relational, and temporal associations.
The system incorporates several self-evolution mechanisms that operate after initial deployment. MindMemEvolve applies validation-driven evolutionary search to refine memory schemas, using error-informed mutation and crossover to discover first-order and higher-order properties relevant to a target domain. During offline periods, a dreaming process consolidates stored records by merging duplicates, resolving conflicts, and preserving provenance. User corrective signals, whether explicit or implicit, further identify and revise inaccurate or misaligned entries without conflating transient task feedback with durable memory updates.
A separate component, MindSkillEvolve, converts accumulated agent execution trajectories into versioned, reusable skills through unsupervised or score-guided refinement. Together these mechanisms address the rigidity of earlier memory systems, which typically fix extraction policies and organization strategies at design time.
On standard benchmarks MindMemOS records 94.03 percent accuracy on LOCOMO and 70.63 percent on PersonaMem. The skill-evolution module additionally raises success rates on SpreadsheetBench by 9.2 percentage points relative to an initial-skill baseline. The architecture exposes these capabilities through service abstractions that integrate with diverse agent frameworks without requiring changes to the underlying foundation models.
Why it matters
This paper provides advanced AI researchers with a rigorous framework for solving long-term memory and skill evolution in LLM agents. Its structured approach to memory consolidation and feedback aligns with the Dutch AI ecosystem's drive toward robust, transparent, and highly capable autonomous systems.
