fuyuxiang/echo-agent

What it solves

Echo Agent is a self-hosted, long-running AI agent designed to move beyond one-off Q&A. It addresses the problem of "memory bloat" in long-term interactions by implementing a structured memory system and allows the agent to evolve its skills based on real-world execution trajectories rather than remaining static after deployment.

How it works

The system operates on an agent loop (Event $\rightarrow$ Context $\rightarrow$ Model $\rightarrow$ Tool) and utilizes a four-layer cognitive memory structure (Working, Episodic, Semantic, and Archival) with automatic decay and contradiction detection. It features a self-evolution engine that generates candidate improvements from execution logs, which are only promoted after evaluation. To manage large tool outputs, it uses an "output spill" mechanism that saves long results to disk and provides the model with a preview and a retrieval path.

Who it’s for

It is intended for users who want a private, auditable AI agent running on their own servers or local machines that can maintain long-term preferences and experience across multiple sessions and communication channels.

Highlights

• Multi-channel Integration: Supports 14 different channels including Telegram, Discord, Slack, WeChat, and Email, all sharing a single state.
• Self-Evolving Skills: Improves its own capabilities through a trajectory-based improvement and evaluation cycle.
• Cognitive Memory: A multi-layered memory system that prevents memory inflation and ensures continuity across sessions.
• Security-First Design: Includes encrypted credential storage, mandatory approval for high-risk tool calls, and a built-in Web Dashboard for monitoring costs and status.
• Extensible Architecture: Supports MCP (Model Context Protocol) clients, A2A JSON-RPC endpoints, and a plugin system.