convez.ai
SkillAutomaçãoIntermediário
samzong

Agent Brains

samzong/agent-brains

Contexto persistente e reutilizável para agentes de IA especializados

Ver no GitHub
InicialPluginScore · de 0 a 100

Catálogo portável de cérebros de agentes para agentes de codificação (Claude, Cursor, etc.). Organiza identidade, memória durable, skills e workflows em estruturas reutilizáveis que mantêm contexto e conhecimento entre sessões, evitando recomeçar do zero.

Instalação

$npx convez-ai samzong/agent-brains

1 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

Estrelas
9
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Manter continuidade de contexto em agentes de codificação entre sessões
  • Organizar conhecimento de domínio específico (identity, memory, SOPs)
  • Compartilhar brains entre equipes e ferramentas (Claude, Codex, Cursor)

Veredito

Projeto bem pensado para quem trabalha com agentes de IA e precisa de persistência de contexto. A arquitetura (AGENTS.md, SOUL.md, MEMORY.md) é elegante, mas requer disciplina para manter brains coesos—vale a pena para fluxos de trabalho recorrentes em domínios específicos.

README

A tradução para o português deste README ainda está na fila.

agent-brains

Portable domain agent brains for coding agents (Codex, Claude Code, Cursor, and others that read markdown context).

An agent brain is a long-lived context pack for a recurring domain. It keeps identity, judgment criteria, style, durable memory, domain methods, and stable SOPs in one loadable structure so the agent does not start from a blank context every session.

See docs/AGENT_BRAIN_SPEC.md for the full specification.

Skill vs workflow vs brain (see spec for detail):

  • skill — reusable method or tool capability (how)
  • workflow — end-to-end SOP for stable business behavior (what process)
  • agent brain — identity, memory, and domain judgment (who, on what history and standards)

Create brains

NamePurpose
agent-creatorCreate, audit, and evolve agent brains

Bundled skills

NamePurpose
agent-loaderList and load repo-scoped and user-scoped agent brains

Built-in loader: skills/agent-loader/. Copy or symlink it into your agent client's global skills directory to use it outside this repo.

Discovery checks repo-local .agents/agents directories from the current working directory up to the repository root, then $HOME/.agents/agents. It reads only AGENTS.md frontmatter and reports name, description, scope, and path. On load, agent-loader resolves the agent path, then reads AGENTS.md, SOUL.md, and MEMORY.md first. memories/, skills/, and workflows/ are lazy-loaded only when the task needs them or core files point to them.

Optional local routing

For Codex users, codex-agents-local can add a private AGENTS.local.md prompt that nudges Codex to choose an agent brain only when useful:

When a task is domain-specific or recurring enough to benefit from a personal agent brain, proactively use `agent-loader` to list/resolve candidates, load the single best-matching agent core before working, ask only if the match is ambiguous, and keep skills/workflows/memories lazy-loaded.

Brain structure

Each brain lives under agents/:

agents/
└── agent-name/
    ├── AGENTS.md
    ├── SOUL.md
    ├── MEMORY.md
    ├── memories/
    ├── skills/
    │   └── SKILL_NAME/
    │       └── SKILL.md
    └── workflows/
        └── WORKFLOW_NAME/
            └── WORKFLOW.md
  • AGENTS.md — identity, responsibilities, rules, boundaries (name, description frontmatter required)
  • SOUL.md — style, judgment, taste, and tone
  • MEMORY.md — durable main memory
  • memories/ — optional focused sub-memories (by date or topic)
  • skills/<skill-name>/SKILL.md — optional reusable methods and tools
  • workflows/<workflow-name>/WORKFLOW.md — optional end-to-end SOPs

Keep brains narrow. A useful brain knows one domain well instead of trying to manage every task.

Related projects

  • AGENTS.md — open format for agent-facing project instructions
  • OpenClaw workspace — runtime-injected workspace files (similar file semantics, different loading model)

License

MIT

Ver o original no GitHub

Bom para

Relacionados