Raven: a self-improving agent harness with durable memory and proactive event-driven capabilities

What it solves

Raven addresses the limitations of typical "LLM + tools" agent loops, such as context overflow in long sessions, repetitive system prompts, passive waiting for user input, and the loss of useful workflows that remain trapped in chat history.

How it works

Raven acts as a self-improving agent harness built on EverOS. It manages the agent's environment through a "Spine" architecture that handles input and output. It uses a Curator context engine to manage token budgets and a Memory Engine (via EverOS) for durable user and agent memory. It includes a Sentinel proactive engine that allows agents to notice events and take action without being prompted, and SkillForge to turn repeated workflows into reusable, evolving skills.

Who it’s for

It is designed for developers and power users who need agents for terminal-native daily work, custom digital workers, and scenarios where static tool loops are insufficient.

Highlights

• Self-Improving Skills: Detects reusable procedures and evolves them based on execution feedback.
• Proactive Agency: Uses a Sentinel and scheduler to initiate actions and "nudges" based on events.
• Terminal-Native Interface: Features a native TUI (built with React/Ink) and CLI for direct interaction.
• Extensive Messaging Gateways: Ships with 12 adapters for platforms like Telegram, Slack, Discord, and WhatsApp.
• Automated Self-Evolution: Includes raven.evolver to diagnose failing trajectories and apply harness patches via git commits.
• Durable Memory: Integrates with EverOS for long-term user and agent memory across sessions.