Echo Agent is a self-hosted agent designed to run continuously rather than answer one question at a time. It is deployed on your own machine or server and brings models, tools, memory, permissions, and message entry points together in a single process, so conversations, preferences, and task experience carry across sessions instead of starting from zero each time.
Memory is organised in four cognitive layers — working, episodic, semantic, and archival — combined with decay, contradiction detection, and importance re-ranking to keep long-running memory from growing without bound. Retrieval is hybrid: BM25 and FAISS vector recall are fused with weights adapted to the query, and the system degrades to lexical search when FAISS is unavailable. A self-evolution engine records execution traces, generates candidate improvements, and promotes them only after evaluation, with cooldown periods and one-command rollback.
Every entry point shares one agent loop — receive an event, build context, call the model, run tools — and one body of state. Entry points include a CLI, a gateway API, webhooks, cron jobs, and 14 messaging channels covering Telegram, Discord, Slack, WeChat, WeCom, Feishu, DingTalk, QQ, WhatsApp, email, and Matrix. The gateway can be registered as a user-level systemd or LaunchAgent service so it restarts on boot and after a crash, and it listens only on loopback; browser requests carrying a cross-site Origin are rejected.
High-risk tool calls such as command execution and file operations pass through an approval step with three policies (manual, smart, and off), and unattended channels reject risky calls by default. Credentials are encrypted at rest, sessions, memory, and traces stay in the local workspace, and execution logs are auditable. Model routing is per role: main inference, context compression, embedding, and risk approval each pick their own provider and model. Tool output above a configurable limit (6,000 characters by default) is written to disk and replaced with a head, tail, and path that the model retrieves through a read_spill call.
It requires Python 3.11 or later and at least one model API key, and includes an interactive setup wizard, a cost attribution report, a built-in web dashboard, an MCP client with OAuth and dynamic tool registration, and an inbound A2A JSON-RPC task endpoint. The current 0.3.x line is beta, with migrations provided for data-structure changes.
Features
- Four-layer memory: working, episodic, semantic, and archival memory with decay, contradiction detection, and importance re-ranking
- Hybrid retrieval: BM25 and FAISS vector recall fused with query-adaptive weights, degrading gracefully when FAISS is missing
- Skill evolution: trace capture, candidate generation, evaluation against a baseline, and promotion or rejection with rollback
- 14 channels: CLI, gateway, webhook, and cron alongside Telegram, Discord, Slack, WeChat, WeCom, Feishu, DingTalk, QQ, WhatsApp, email, and Matrix, all sharing one state
- Tool approval: manual, smart, and off policies, with unattended channels refusing high-risk calls
- Model routing: separate provider and model choices for inference, context compression, embeddings, and risk approval
- Interoperability: an MCP client with OAuth and dynamic tool registration, plus an inbound A2A JSON-RPC endpoint
- Background service: user-level systemd or LaunchAgent registration with start, status, logs, restart, and uninstall subcommands
- Output spill: oversized tool output is written to disk and retrieved by character range or regular expression through read_spill
- Local-first storage: sessions, memory, traces, and encrypted credentials stay in the workspace directory
- Scheduled jobs: a built-in cron scheduler triggers agent runs on a schedule
- Dashboard and cost: a web panel for conversations, spend, and run status, plus a cost attribution report