Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.argentos.ai/llms.txt

Use this file to discover all available pages before exploring further.

​
General

ArgentOS is an always-on personal AI operating system. It runs a persistent AI agent that you can talk to through Telegram, Discord, WhatsApp, Slack, Signal, iMessage, or a web dashboard. The agent has persistent memory, can execute commands, browse the web, manage tasks, and learn from interactions.
ArgentOS is self-hosted and always-on. Unlike web-based AI chat interfaces:
  • Your agent runs on your hardware with your API keys
  • It has persistent memory that spans all conversations
  • It can execute commands and interact with your system
  • It connects to your existing messaging platforms
  • It can work autonomously through the task and heartbeat systems
ArgentOS has a free Community tier with core functionality. Pro and Enterprise tiers unlock additional features. You still need your own API keys for model providers (Anthropic, etc.).
At minimum, an Anthropic API key or Max subscription setup token. Optionally, you can add keys for MiniMax, Z.AI, OpenRouter, Ollama (free, local), and ElevenLabs (for TTS).

​
Setup

Node.js 22+ is required. Use nvm (not Homebrew) on macOS to avoid ABI compatibility issues with native modules like better-sqlite3.
nvm install 22
nvm use 22
Yes, through WSL2 (Windows Subsystem for Linux). Install Ubuntu in WSL2 and follow the Linux setup guide.
Yes, ArgentOS supports ARM64 Linux. A Raspberry Pi 4+ with 4GB+ RAM is recommended.

​
Usage

Yes. ArgentOS supports running all channels simultaneously. Your agent shares the same memory and context across all platforms.
Yes. MemU (the memory system) persists information across all sessions and channels. The agent stores important facts, preferences, and events automatically.
The model router scores each message’s complexity and routes simple queries to cheap models (Haiku) and complex tasks to powerful models (Opus). Most conversations use the balanced tier (Sonnet), and simple greetings use the local tier (free).
Yes. The heartbeat system periodically wakes the agent to check for pending tasks. Combined with quiet hours configuration, the agent can work through task queues autonomously.

​
Security

Only the messages sent to the LLM API (Anthropic, etc.) leave your machine. Memory, tasks, sessions, and configuration are all stored locally. ArgentOS does not phone home or collect telemetry.
Only if configured to. The sandboxing system controls what the agent can do. In sandboxed mode (default), dangerous operations require your approval. In locked mode, command execution is disabled entirely.
In sandboxed mode, destructive commands require your approval. The agent cannot delete files, modify permissions, or run sudo without explicit confirmation.

​
Technical

Two separate systems:
  1. MemU — The active agent memory at ~/.argentos/memory.db. This is the agent’s real brain, with semantic search, emotional context, entity tracking, and hybrid vector reranking.
  2. claude-mem — Used by Claude Code sessions only. Completely separate from the ArgentOS agent.
If you see references to openclaw-mem, that was a deprecated legacy plugin from early development. It is no longer maintained and has been fully superseded by MemU.
ArgentOS was inspired by OpenClaw’s original vision of a self-hosted AI agent but is an independently developed AI operating system. ArgentOS has 90,000+ lines of original code powering systems with no OpenClaw equivalent, including persistent memory (MemU), consciousness kernel (contemplation + SIS), voice presence (AEVP), intent governance, knowledge library, execution worker, and more. A small runtime compatibility layer (@mariozechner/pi-* packages) remains from the shared heritage and is being actively replaced by argent-native code.
The base installation uses under 100MB. Memory and session data grow over time. A typical installation with months of use is under 500MB.