The agent harness that treats memory and scaffolding as the product

What it does

Raven is a Python-based runtime that sits between you and a collection of LLM providers. It runs as a terminal TUI, a CLI, or a messaging gateway across twelve platforms, and uses its underlying EverOS layer to persist user memory, agent memory, and world knowledge across sessions. The goal is to turn repeated workflows into evolving skills rather than letting them vanish into old chat transcripts.

The interesting bit

Instead of treating the agent loop as the whole product, Raven treats the surrounding infrastructure—context budgets, proactive nudges, skill materialization, and sandboxed execution—as the actual deliverable. The TUI is a React/Ink application talking to the Python backend over typed RPC, which is an unusually thoughtful bit of terminal UI architecture for a Python project.

Key highlights

• Memory-first design: separates user memory, agent memory, and world knowledge via EverOS, making sessions resumable and patterns reusable.
• SkillForge: detects reusable workflows, writes them to skill files, and tracks feedback to evolve instructions over time.
• Proactive engine: Sentinel watches events and evaluates whether to nudge the user, rather than waiting passively for the next prompt.
• Broad provider and gateway support: plugs into OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, Copilot, and bridges to Telegram, Slack, Discord, WhatsApp, Matrix, and others.
• Explicit context management: uses “Curator” and legacy engines with token budgets and fail-safes instead of silently truncating history.

Verdict

Raven is worth a look if you want a persistent, terminal-native agent runtime that can bridge chat apps and remembers context across sessions. If you just need a quick OpenAI CLI wrapper or a static prompt chain, this is likely more harness than you need.