EverMind-AI/Raven

What it solves

Standard AI agent tools often struggle with long sessions that overflow context, passive behavior that requires constant user prompting, and workflows that remain trapped in chat history rather than becoming reusable assets. Raven addresses these issues by providing a robust harness that manages memory, context, and skill evolution.

How it works

Raven acts as a sophisticated layer around LLMs, utilizing the EverOS memory substrate to maintain durable user and agent memory across sessions. It employs a "Spine" architecture to manage the agent loop, featuring a context engine to prevent context collapse, a proactive engine (Sentinel) to allow agents to initiate actions, and a self-evolution pipeline (Evolver) that uses benchmarks to automatically improve the harness via git commits.

Who it’s for

• Developers looking for a terminal-native runtime and agent templates to build digital workers.
• Power users who need an agent capable of deep research and proactive task management.
• AI researchers interested in self-improving agent architectures and benchmark-driven evolution.

Highlights

• Self-Improving Skills: Automatically detects and materializes reusable workflows into evolving skills.
• Durable Memory: Uses EverOS to separate user, agent, and world knowledge for long-term persistence.
• Proactive Agency: Includes a scheduler and nudge policy so agents can act without waiting for user input.
• Deep Research: An opt-in tool for multi-source investigation via MiroThinker.
• Extensive Connectivity: Supports 12 messaging gateways including Telegram, Slack, Discord, and WhatsApp.