EverMind has released EverOS, an open-source memory runtime for AI agents, available under an Apache 2.0 license. It addresses a fundamental challenge faced by agent developers: large language models are inherently stateless. Once a conversation ends, the context is lost.
EverOS introduces a novel approach. Instead of storing memory solely within a vector database, it persists memory as plain Markdown files. These files serve as the authoritative source of truth that agents can read, edit, and search across sessions. This design enables a more transparent and flexible memory system.
Key Features
- Markdown-First Storage: Memory is stored in human-readable Markdown files, making it easy to inspect, edit, and version control.
- Hybrid Retrieval (BM25 + Vector): Combines the precision of keyword-based BM25 retrieval with the semantic understanding of vector embeddings for more accurate context retrieval.
- Self-Evolving Skills: Agents can autonomously refine and expand their skills over time based on accumulated memory and experience.
- Open Source & Extensible: Built as an open-source runtime that can be integrated into any agent architecture, encouraging community contributions and customization.
Why EverOS in 2026?
As AI agents become more autonomous and long-running, persistent memory has become a critical bottleneck. Traditional vector databases alone often fail to capture nuanced, evolving contexts. EverOS addresses this by combining multiple retrieval strategies and leveraging Markdown as a universal, durable format. This aligns with the growing trend of transparent, verifiable agent systems that prioritize developer control and data portability.
Getting Started
EverOS is designed to be lightweight and modular. Developers can clone the repository, link it to their agent framework, and begin persisting and querying memory with minimal configuration. The project welcome contributions, and detailed documentation is available on the official GitHub repository.
via MarkTechPost