Echo Agent Documentation¶
Echo Agent is a self-hosted, long-running AI Agent runtime with persistent memory, skills evolution, and multi-channel integrations.
Quick Navigation¶
-
:material-rocket-launch:{ .lg .middle } Getting Started
Install and have your first conversation in 5 minutes
-
:material-server-network:{ .lg .middle } Background Deployment
Run as a daemon or systemd service for 24/7 availability
-
:material-chat-processing:{ .lg .middle } Platform Integrations
Connect your Agent to DingTalk, Feishu, WeChat, Slack, Telegram, and 14 channels total
-
:material-puzzle:{ .lg .middle } Extend & Develop
Write custom skills, plugins, and channel adapters
Core Capabilities¶
| Module | Description | Details |
|---|---|---|
| Agent Loop | Receive events → build context → call model → execute tools, shared across all entry points | Agent loop |
| Cognitive Memory | Working / Episodic / Semantic / Archival four tiers, with decay, contradiction detection and importance reranking | Memory system |
| Hybrid Retrieval | BM25 + FAISS vector fusion, adaptive weighting per query, graceful degradation without FAISS | Knowledge base |
| Self-Evolution | Trajectory recording → candidate generation → eval comparison → promote/reject, with cooldown and rollback | Evolution & evaluation |
| Model Routing | Main reasoning, context compression, embeddings and risk approval each configurable with independent provider and model | Routing & fallback |
| Tool Approval | Three modes: manual / smart / off, unattended channels default to denying high-risk calls |
Tools & permissions |
| Multi-model support | OpenAI, Anthropic, Gemini, Bedrock, OpenRouter, plus OpenAI-compatible endpoints (DeepSeek, Qwen, Kimi, GLM, Ollama) | Provider overview |
| 14 channel adapters | CLI, Cron, DingTalk, Discord, Email, Feishu, Matrix, QQ Bot, Slack, Telegram, Webhook, WeCom, WeChat, WhatsApp | Channel setup |
| Cross-Process Interop | Inbound A2A JSON-RPC service + MCP client (with OAuth), with dynamic tool registration | MCP · A2A |
| Plugin system | Register external plugins via entry-points | Using plugins |
| Dashboard | Built-in web panel for conversations, cost and runtime status | Dashboard |
| Scheduled tasks | Built-in cron scheduler for time-triggered Agent execution | Scheduled jobs |
| Output Preservation | Oversized tool output is spilled to disk; the model sees a head/tail preview plus a retrieval path and can pull the full text back with read_spill |
Context compression & spill |
| Local-First | Sessions, memory, traces and credentials stored in the workspace by default, credentials encrypted at rest | Security model |
Project Status¶
Beta Stage
Echo Agent is currently at v0.3.8 in Beta. The core API is stabilizing, but breaking changes may still occur in:
- Configuration file format (
config.yamlschema) - Plugin / skill API interfaces
- Database schema (migration provided via
echo-agent migrate)
Review the CHANGELOG and back up your data before upgrading.
System Requirements¶
- Python 3.11+
- Linux / macOS / Windows (WSL2 recommended)
- At least one model API key (OpenAI, Anthropic, etc.)
License¶
Echo Agent is open-source under the MIT License.