An infrastructure layer for AI agents that refuse to forget

What it does

Zylos is a Node.js-based life-support system for AI agents. It wraps Claude Code or OpenAI Codex in persistent infrastructure: a five-tier memory stack, an autonomous scheduler, multi-channel messaging through Telegram and Lark, and a self-healing activity monitor. Everything runs on your own server, with messages audited to SQLite and a web console for direct access.

The interesting bit

The project treats session amnesia as the central enemy. Most frameworks isolate conversations by channel; Zylos routes every message through a single C4 communication bridge so the agent has one identity everywhere. It also auto-saves memory at 75% context capacity before compaction hits, which is the kind of defensive programming usually reserved for database transaction logs.

Key highlights

• One continuous agent across Telegram, Lark, and a web console via a unified C4 gateway backed by SQLite audit trails
• Five-layer “Inside Out” memory architecture (identity → state → references → sessions → archive) that survives restarts
• Built-in scheduler and activity monitor with crash recovery, heartbeat checks, and PM2-based auto-recovery
• Swappable AI runtime: switch between Claude Code and Codex without losing memory or skills
• Runs on a flat-rate Claude subscription rather than per-token billing; the README notes community reports of $500–$3,600 monthly bills from token-based alternatives

Caveats

• Requires a Linux or macOS server and Node.js 20+; Windows users must install via SSH to a remote machine
• The “self-evolving” capability is really just the underlying runtime writing new code, not a separate learning system

Verdict

Worth a look if you need a single, always-on agent that maintains context across multiple chat platforms. Skip it if you want a lightweight, ephemeral chatbot or native Windows support.