ARIA - Local AI Agent

A personal AI assistant that runs entirely on your machine via Ollama. Persistent memory, autonomous skill creation, self-evaluation, agentic tool use - no API keys, no cloud, no cost.

---

Requirements

• Python 3.10+
• Ollama running locally

---

Setup

1. Install Ollama

# macOS / Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows - download installer from https://ollama.com

2. Pull a model

ollama pull llama3.2        # 2GB, fast - good starting point
ollama pull mistral         # 4GB, better quality
ollama pull qwen2.5:7b      # excellent reasoning

3. Run ARIA

python3 server.py

Open http://localhost:7842

---

UI Overview

Three-panel layout:

Left sidebar - 7 tabs (Config / Memory / Skills / Code / KB / Sessions / Cron)

• Config - Ollama URL, model selector with live readiness check, quick actions, stats
• Memory - view and edit agent memory (facts) and user profile
• Skills - browse, create, and edit procedural skills
• Code - search Stack Overflow, MDN, DevDocs, GitHub, PyPI, npm
• KB - full-text + vector knowledge base, paste notes and docs
• Sessions - browse and search past conversations
• Cron - schedule recurring tasks

Center - chat interface with tool log per message

Right sidebar - SOUL.md viewer (click to edit), evolution log

---

How the agentic loop works

When you send a message, ARIA runs a multi-step loop before replying:

1. Searches memory and knowledge base for relevant context
2. Calls web_search or wikipedia if it needs new information
3. Runs execute_code for Python/Bash when needed
4. Saves important facts with memory(add)
5. Creates a skill with skill_create after solving complex tasks
6. Returns the final answer

Tool calls are visible - click "Tool log" on any AI response to inspect every step.

---

Slash commands

/skills          - list all available skills
/skill-name      - invoke a specific skill by name

---

Data

Everything lives in ~/.aria/ - nothing leaves your machine:

~/.aria/
├── state.db              ← SQLite: sessions, messages (FTS5), knowledge, cron jobs
├── SOUL.md               ← Evolving system personality (editable in the UI)
├── evolution.json        ← Generation history and self-evaluation scores
├── memories/
│   ├── MEMORY.md         ← Agent facts (2200 char cap, § delimited)
│   └── USER.md           ← User profile and preferences (1375 char cap)
└── skills/
    └── skill-name/
        └── SKILL.md      ← Agent-created procedural memory

---

Environment variables

ARIA_PORT=8080 python3 server.py

---

Feature comparison

Feature	Hermes Agent	ARIA
Local LLM (Ollama)	✅	✅
Live model readiness check	-	✅ /api/ps + generate ping
Dual memory (facts + user profile)	✅	✅
Memory capacity management	✅ 2200+1375 chars	✅ same limits
Skills system (procedural memory)	✅	✅
Agent creates skills autonomously	✅	✅
Skill create/edit/delete in UI	✅	✅
SQLite session storage (FTS5)	✅	✅
Session search across history	✅	✅
Multi-step agentic tool loop	✅	✅
Code execution (Python/Bash)	✅	✅
Web search (DuckDuckGo)	✅	✅
Wikipedia research	✅	✅
File read/write	✅	✅
RAG (vector + FTS hybrid)	✅	✅
Dev resource search (SO, MDN, etc.)	-	✅
Self-evaluation & SOUL.md rewriting	✅	✅
Scheduled tasks (cron)	✅	✅
SOUL.md personality file	✅	✅
Live soul editor in UI	-	✅
Messaging platforms	✅ 20+	❌ browser only
Voice mode	✅	❌
Sub-agents	✅	❌
Docker/SSH backends	✅	❌
Skills Hub (community)	✅	❌