AgentOS

AgentOS is an open-source TypeScript framework for building autonomous AI agents for agentic orchestration. It provides a runtime for agents that can maintain long-term memory, adapt their behavior, and generate their own tools during execution. The software is deployed as a server-side library and can be integrated into web applications, CLI tools, or background daemons.

The framework focuses on persistent cognitive memory and multi-agent collaboration. It allows developers to define agents with specific personalities and instructions, then organize them into teams using various coordination strategies. It supports a wide range of integrations, including multiple LLM providers and various communication channels such as Slack, Discord, and Telegram, making it suitable for complex automated workflows.

Key features

• Runtime tool forging that allows agents to write and execute TypeScript functions in a sandbox
• Cognitive memory system using eight neuroscience-backed mechanisms like Ebbinghaus decay
• Support for 11 LLM providers including OpenAI, Anthropic, Gemini, and local options via Ollama
• Optional HEXACO personality vectors to bias retrieval and decision making
• Multi-agent orchestration with six strategies including sequential, debate, and hierarchical
• Soul files for managing agent identity and long-term memory via a markdown wiki
• RAG pipeline with seven vector backends and support for GraphRAG and HyDE
• Built-in guardrails for PII filtering, code safety, and content policy

AgentOS is designed for developers building complex AI systems that require state persistence across long sessions. Its architecture separates the kernel logic from the prompt, ensuring that personality traits and memory remain stable even under heavy context pressure. The system includes a dedicated benchmark harness for measuring memory retrieval accuracy, specifically targeting LongMemEval-S and LongMemEval-M benchmarks. It also features a registry for sharing curated skills and extensions, allowing agents to import pre-defined capabilities at startup.

It is a high-performance orchestration layer for autonomous agents with a specific emphasis on reproducible memory benchmarks and emergent tool creation.