Skip to main content
The runtime process executes autonomous behavior in each Ara sandbox.
Runtime operations visual
Runtime loop: tools, memory, and delegation running in one isolated session.

​
Responsibilities

  • Tool-loop orchestration
  • Skill resolution and loading
  • Memory read/write in markdown files
  • Channel ingestion and outbound delivery
  • Subagent and spawn-based delegation

​
Runtime configuration

Primary config file: 
/root/.ara/runtime-config.json
Cloud control plane updates this file when users change providers, channels, or integrations.

​
Skills hierarchy

The runtime resolves skills in layered precedence:
  1. workspace/skills
  2. self-skills (agent-created)
  3. global user skills
  4. builtin skill bundles
Each skill is a directory containing SKILL.md.

Skills system

See discovery, installation, and agent-created skill workflows.

​
Tool categories

CategoryExamples
Filesystemread/write/edit/search/list
Executionshell/exec with policy constraints
Webbrowser automation, web fetch/search
MCPexternal tool servers and discovery
Skillssearch/install/manage/create
Delegationspawn, subagent

​
Memory model

Runtime memory is markdown-centric:
  • MEMORY.md for core persistent memory
  • Daily notes partitioned by date
  • Prompt context can freeze memory snapshots per session for stable caching

​
Channels and gateway

The runtime gateway receives websocket frames proxied by cloud API. Channel adapters process provider-specific payloads and normalize them into runtime message events.

​
Why this matters

  • Gives agents local context and continuity
  • Enables recursive improvement via skill and memory writes
  • Supports the same autonomy model across app chat, APIs, and external messaging channels