PAI Universal
starlink-awaken/pai-universal
Infrastructure agnóstica de IA pessoal com sistema de memória e algoritmo universal
Ver no GitHubFramework agnóstico de infraestrutura pessoal de IA que qualquer agente pode usar. Implementa um motor universal de tarefas (Current State → Ideal State), sistema de memória em três camadas, gerenciamento de identidade e lifecycles de hooks. Totalmente desacoplado de plataformas específicas via camada de adaptadores.
Instalação
npx convez-ai starlink-awaken/pai-universal1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Criar assistentes de IA customizados que funcionam em múltiplas plataformas (Claude, Gemini, APIs custom)
- Implementar sistemas de memória persistente (WORK → LEARNING → KNOWLEDGE) para agentes
- Construir agentes com identidade clara e objetivos de vida estruturados (Telos, Missions, Strategies)
Veredito

PAI Universal é uma arquitetura sofisticada e bem pensada para quem quer construir agentes de IA sérios com persistência, memória e portabilidade entre plataformas. Exige typescript e conhecimento arquitetural, mas oferece abstrações elegantes e reutilizáveis para qualquer cenário de automação inteligente.
README
A tradução para o português deste README ainda está na fila.
PAI Universal
A platform-agnostic Personal AI Infrastructure framework. Any AI agent can use it.
What This Is
PAI Universal extracts the core ideas from PAI v5.0 into a portable framework:
- Algorithm — universal task engine: Current State → Ideal State via verifiable iteration
- ISA — Ideal State Articulation, the single source of truth for any task
- Memory — three-tier compounding system: WORK → LEARNING → KNOWLEDGE
- Telos — life goals, missions, strategies, challenges
- Identity — who you are and who your DA (Digital Assistant) is
- Skills — self-activating composable domain units
- Hooks — lifecycle event system for any platform
The framework is not tied to any specific AI platform. It defines adapter interfaces for tools, hooks, agents, config, and skills. Platform plugins implement these interfaces.
Architecture
┌─────────────────────────────────────────────┐
│ Your AI Agent │
│ (Claude Code, Gemini CLI, etc.) │
└──────────────┬──────────────────────────────┘
│ uses
┌──────────────▼──────────────────────────────┐
│ PAI Universal Core │
│ ┌─────────┬─────────┬──────────┬────────┐ │
│ │Algorithm│ ISA │ Memory │ Telos │ │
│ ├─────────┼─────────┼──────────┼────────┤ │
│ │Identity │ Skills │ Hooks │Context │ │
│ └─────────┴─────────┴──────────┴────────┘ │
├─────────────────────────────────────────────┤
│ Adapter Layer │
│ ┌──────────┬──────────┬────────┬────────┐ │
│ │ Tools │ Hooks │ Agents │ Config │ │
│ └──────────┴──────────┴────────┴────────┘ │
├─────────────────────────────────────────────┤
│ Platform Plugins │
│ Claude Code │ Gemini CLI │ Custom API │ … │
└─────────────────────────────────────────────┘
Quick Start
# Clone and install
cd ~/ToolBox
git clone <your-repo> pai-universal
cd pai-universal
bun install
# Type check
bun run typecheck
Use in Code
import { createPAI } from './src/index.ts';
// Write your own plugin or use a built-in one
import { claudeCodePlugin } from './plugins/claude-code/index.ts';
const pai = await createPAI(claudeCodePlugin(), './workspace');
// Access core systems
const state = pai.algorithm.getState();
const memory = await pai.memory.loadAll();
const context = pai.context.build(fullContext);
Write Your Own Plugin
import type { PAIPlugin, PAIAdapter, PAIContext } from 'pai-universal';
import { BasePAIAdapter } from 'pai-universal';
class MyAdapter extends BasePAIAdapter {
readonly name = 'my-platform';
readonly version = '1.0.0';
readonly description = 'My platform adapter';
}
export function myPlugin(): PAIPlugin {
return {
name: 'my-platform',
version: '1.0.0',
async createAdapter(context: PAIContext): Promise<PAIAdapter> {
return new MyAdapter();
},
};
}
See plugins/README.md for the full adapter development guide.
Project Structure
src/
├── types.ts # All core type definitions
├── adapter.ts # Base adapter implementations
├── registry.ts # Plugin registry
├── algorithm.ts # Algorithm engine + ISA manager
├── memory.ts # Three-tier memory system
├── identity.ts # Principal + DA identity management
├── telos.ts # Life goals management
├── skill.ts # Skill discovery and registry
├── context-builder.ts # Assembles context into prompts
└── index.ts # Public API + createPAI helper
templates/ # Platform-agnostic templates
├── algorithm/ # Algorithm v6.3.0 definition
├── isa/ # ISA format specification
├── identity/ # Principal + DA identity templates
├── telos/ # Mission, goals, challenges, etc.
├── memory/ # Memory system structure
├── prompts/ # System prompt + mode templates
└── config/ # Default configuration
plugins/ # Platform adapter plugins
├── claude-code/ # Claude Code CLI adapter
└── README.md # Plugin development guide
Core Concepts
Algorithm (v6.3.0)
The universal task engine. Every task follows: Current State → Ideal State via verifiable iteration (ISC).
Seven phases: OBSERVE → THINK → PLAN → BUILD → EXECUTE → VERIFY → LEARN
Five effort tiers (E1-E5) with escalating requirements for ISC count, thinking depth, and delegation.
ISA (Ideal State Articulation)
A structured document that is simultaneously:
- The ideal state description
- The test harness
- The build verification
- The done condition
- The system of record
12 sections, fixed order, tier-gated requirements.
Memory
Three tiers that compound over time:
- WORK — active work, session-scoped
- LEARNING — extracted patterns, failure modes, corrections
- KNOWLEDGE — stable, permanent, rarely pruned
Telos
Your life operating parameters:
- Missions (durable north stars)
- Goals (measurable, ISC-bearing)
- Strategies, Narratives, Challenges, Core Models
Identity
Two layers:
- Principal — who you are, career, worldview, preferences
- DA — your Digital Assistant's name, personality, writing style, autonomy boundaries
Adapter Interfaces
| Interface | Purpose | Key Methods |
|---|---|---|
ToolAdapter | File ops, commands, search | readFile, writeFile, runCommand, search |
HookAdapter | Lifecycle events | on, emit, off |
AgentAdapter | Agent dispatch | dispatch, register, list |
ConfigAdapter | Configuration | get, set, onChange |
SkillAdapter | Skill management | load, execute, register |
License
MIT
Credits
Ported from PAI v5.0 by Daniel Miessler. Core concepts preserved; Claude Code dependencies abstracted into adapter layer.