ArX

AI-Powered Creative Engine

View on GitHub

ArX is Arvolve's proprietary AI engine: a modular, production-ready system that turbocharges creative workflows through autonomous AI agents, persistent memory, and comprehensive tool integration.

Combines text generation, vision analysis, image/video/3D creation, research, and autonomous proactive operation into a unified system. Built for real-world production use in content creation, research, and workflow automation.

System Architecture

ArX Architecture Graph

Core Architecture — 6 Layers

I — Integrations

src/I_integrations/

API wrappers for OpenAI, Ollama, Replicate, Google, web crawling

II — TextGen

src/II_textGen/

Core hub: memory, RAG, tool orchestration, structured output

III — Agents

src/III_agents/

Multi-agent system + shared memory + autonomous heartbeat

IV — Pipelines

src/IV_pipelines/

Domain workflows (image, video, 3D, research, pitch decks, code factory)

V — Branches

src/V_branches/

Interfaces: FastAPI (83 endpoints), Telegram, Slack, Desktop GUI, Blender

VI — Utils

src/VI_utils/

Utilities: GDrive sync, image/video processing, code tools

Key Features

Memory System

  • · Short-term (STM) — recent conversation with auto-trimming
  • · Long-term (LTM) — semantic memory (facts, preferences, goals, procedures)
  • · Identity — versioned personality/behavior rules with rollback
  • · Journal — scratchpad for notes and observations
  • · Ops Board — outcome-driven task management (NOW/WAITING/UPCOMING/INBOX)

Tool Integration

  • · Research — web crawling, multi-source search (DDG, Wikipedia, Exa)
  • · Generation — images (Replicate, GPT), video (Wan, Veo, Sora), 3D (Tripo, Hunyuan), documents, graphs
  • · Communication — email via Gmail API
  • · Utilities — date/time, file ops, GDrive sync

Heartbeat — Autonomous Operation

  • · Self-editing checklist (HEARTBEAT.md) — AI updates its own operating instructions
  • · Cron scheduling — time-based recurring tasks
  • · Loop prevention — detects and suppresses repeated actions
  • · Ops Board integration — executes tasks from NOW list autonomously
  • · Full tool access — same capabilities as interactive mode

Multi-Agent System

  • · AgentGen — orchestrate multiple specialized agents with isolated memory
  • · Agent routing — LLM-powered task assignment to best-fit agent
  • · Shared memory — cross-agent task continuity via AgentsMemory
  • · Parallel execution — worker threads for concurrent agent operations

RAG with Caching

  • · Separate indices for LTM and session context
  • · Hash-based cache invalidation (only re-index when content changes)
  • · Semantic search with configurable chunk size/overlap

FastAPI Server v0.5.0 · 83 endpoints

/openai 6 Raw OpenAI API access
/text 10 TextGen with memory/RAG
/audio 3 TTS, STT, music generation
/image 4 Generate, upscale, download
/video 2 Video generation (multiple providers)
/threed 2 3D generation (Hunyuan, Trellis2, Tripo)
/tools 3 Web crawl, graphs, datetime
/agents 14 Multi-agent orchestration
/pipelines 17 High-level workflows
/heartbeat 11 Proactive scheduling, cron jobs
/gdrive 5 Background sync to Google Drive

Interfaces

FastAPI

Production REST API — 83 endpoints

Desktop GUI

CustomTK chat with memory inspector + heartbeat controls

Telegram Bot

Conversational AI with mode switching (chat/alin/arvolve)

Slack App

Workspace integration

Blender Addon

In-app AI tools for 3D artists

Tech Stack

LLM Providers

OpenAI (GPT-5.2, GPT-4o, o3-mini) · Ollama (Gemma3, Llama3, Mistral, Phi3)

Generation APIs

Replicate (image, video, 3D) · Tripo (3D from images/text)

Integrations

Google APIs (Drive, Gmail, Maps, Search) · FastAPI + Uvicorn · SQLite · FAISS

Philosophy

First Principles

Strip assumptions, build from solid truths.

Continuous Adaptability

Evolve with new technologies as they emerge.

Human-Centered

AI augments creativity, doesn't replace it.

Timelessness

Build tools that remain relevant beyond current trends.

Less is More

Simple, elegant, effective solutions over complexity.

Project Lifecycle

1

Spark

Raw intuition captured privately. Incomplete by design.

2

Draft

Structured exploration. Key assumptions and frames established.

3

Crystallized

Meets the Definition of Done. Internally coherent and decision-ready.

4

Published

Public layer released to the Atlas. Stable permalink.

5

Restricted

Full bundle available selectively — blueprints, implementation notes.

Crystallized — Definition of Done

  • Thesis clarity — Expressible in one sentence without losing meaning.
  • Constraints stated — Assumptions, tradeoffs, and non-goals explicit.
  • Time horizon labeled — Near / Emerging / Speculative declared and justified.
  • Feasibility gradient — Path from present → plausible execution sketched.
  • Internal consistency — No contradictions across rationale, visuals, and mechanics.
  • Signal metrics — Evaluable using the curation dimensions.
  • Next actions — At least one concrete first step for a builder or researcher.