EverMind-AI/Raven

What it solves

Raven provides a comprehensive system for managing long-running AI agent work. It addresses the difficulty of scaling agents beyond narrow tasks by integrating terminal-first execution, long-term memory, and self-improving capabilities into a single "Agent Harness," reducing the need to manually design complex harnesses for every new model or domain.

How it works

Raven operates as a central loop that coordinates several core systems:

• EverOS Memory: Provides durable user and agent memory across sessions.
• SkillForge: Manages a catalog of skills that can be retrieved, fed back into, and evolved.
• Context Engine: Handles token budgeting and context assembly to keep the most relevant information available to the LLM.
• Proactivity: Uses sentinels and scheduling to allow agents to perform deferred decisions and scheduled work.
• Evolver: Implements reproducible evaluation loops to improve agent procedures.
• Deep Research: Integrates with MiroThinker for open-ended web search and multi-source analysis with inline citations.

Who it’s for

It is designed for developers and AI researchers who need a robust, inspectable environment for deploying autonomous agents that require persistence, complex tool use, and the ability to evolve their performance over time.

Highlights

• Terminal-First Interface: Features a TUI (Terminal User Interface) and a CLI for streamlined setup and operation.
• Local Tracing: Includes a local dashboard to inspect LLM calls, tool usage, and memory recall without sending data to external services.
• Broad Integration: Supports a wide array of LLM providers (OpenAI, Anthropic, DeepSeek, etc.) and messaging gateways (Telegram, Slack, Discord, etc.).
• Self-Improving: Built-in mechanisms for evaluation and skill evolution to improve agent performance over time.