An AI agent that remembers, forgets, and asks for permission

What it does Echo Agent is a self-hosted AI agent framework meant to run persistently on your own hardware. It collapses a dozen entry points—CLI, API gateway, messaging bots, and cron jobs—into one shared execution loop with unified memory and permissions. The goal is continuity: instead of stateless chat, it carries structured context across sessions and refines its behaviors over time.

The interesting bit The memory system treats forgetting as a feature, not a bug. It splits recollection into four layers (working, episodic, semantic, archival), applies decay curves and contradiction detection, and falls back from vector search to keyword retrieval if FAISS is absent. There is also a self-evolution pipeline that generates candidate skill improvements from execution traces, evaluates them, and promotes or rolls them back—an attempt to make the agent learn from use rather than simply log it.

Key highlights

• Four-layer cognitive memory with explicit forgetting curves and contradiction checks to combat context bloat.
• LLM-based approval workflow for high-risk tools like command execution and file operations, with encrypted credential storage and audit logging.
• Native MCP client and A2A JSON-RPC support for plugging into external tool ecosystems and other agents.
• Multi-model routing that lets you assign different providers to inference, embedding, compression, and risk approval tasks.
• Twelve-channel input normalization feeding one shared state, covering CLI, Gateway, Cron, Telegram, Discord, Slack, WeChat, Feishu, and DingTalk.

Caveats

• The gateway binds to loopback by default and rejects cross-origin browser requests to block CSRF; remote access means tunneling or explicitly weakening that boundary.
• The README notes that self-evolved skills are “evaluated” before promotion, but it never explains the evaluation criteria or failure rates.
• FAISS is optional and the system degrades to BM25, though the README doesn’t detail the accuracy or latency trade-offs of that fallback.

Verdict Worth exploring if you need a persistent, self-hosted automation layer that treats memory as an engineering problem rather than a vector dump. Skip it if you just want a quick OpenAI wrapper or have no use for a long-running Python service.