Claudart
vankhaivn/Claudart
A modular and self-evolving rules framework to optimize context windows for Claude and other AI coding agents.
Instalação
npx convez-ai vankhaivn/Claudart1 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
README
A tradução para o português deste README ainda está na fila.
The problem. AI coding agents leak state: every session starts blind, plans die in chat, decisions get re-discovered weekly, and
CLAUDE.md/AGENTS.mdbloats into a token sink nobody trusts.
CLAUDART fixes this with a handful of slash commands over a layered markdown memory model — plain files under .claude/ and .codex/, versioned in git, reviewable in PRs, readable offline. No vector DB, no daemon, no cloud.
Install
# Flag after `bash -s --`: --claude (default) · --codex · --both · --force (overwrite)
curl -fsSL https://raw.githubusercontent.com/vankhaivn/Claudart/main/install.sh | bash -s -- --claude
Already have your own setup, or upgrading? install.sh does a fresh copy that clobbers customizations. Instead, paste this into your agent — it reads the repo, diffs against your project, and merges only what you approve:
Read https://raw.githubusercontent.com/vankhaivn/Claudart/main/INTEGRATE.md and follow it to integrate CLAUDART into this project. Ask me before touching anything I've customized.
What it solves
| Pain | CLAUDART's answer |
|---|---|
| Every session starts blind | /start — reads current state, active tasks, the knowledge index, and recent commits |
| Plans die when the session closes | /plan — a persistent task doc that survives any pause |
| A productive session hits the context ceiling | /handoff — a single-slot baton with the session's reasoning state; the next /start resumes from it |
| Same decisions re-discovered weekly | /learn — graduates recurring patterns into path-scoped rules/ |
| Durable facts have nowhere to live | knowledge/ — descriptive facts (domain, architecture, glossary), surfaced every session |
CLAUDE.md bloats and burns tokens | /refactor-memory — trims it to a lean index; behavior → rules, facts → knowledge |
| Memory drifts silently; agents go off-scope | /doctor flags rot; clean-code-reviewer, security-auditor, and agent-delegation keep work bounded |
The memory model
SESSION STATE (volatile) DURABLE REFERENCE (survives sessions)
CONTEXT.md JOURNAL.md rules/ · guidelines/ knowledge/
what's true now what happened how to behave what the project is
(declarative) (history log) (prescriptive) (descriptive facts)
always loaded never loaded auto-loads on INDEX on /start,
in context (audit only) a matching path detail on demand
/checkpoint rebuilds CONTEXT.md, retires history to JOURNAL.md (audit-only, never loaded), and writes durable facts to knowledge/. /learn promotes recurring behavior to rules/. /doctor and /refactor-memory keep knowledge honest — flagging drift and re-checking each fact against the code.
Quick start
# In a project with CLAUDART installed
/start # orient the session
/plan add JWT middleware # persistent task doc — agent waits for your approval before coding
/handoff # context window nearly full — distill reasoning state, resume via /start
/checkpoint # rebuild CONTEXT.md + sync state at session end
/learn # graduate recurring decisions into rules
/doctor # health check when the setup feels off
Codex CLI: same flow, swap / for $codex- (e.g. $codex-start).
Documentation
docs/WORKFLOW.md is the manual — architecture, full task lifecycle, every command, directory layout. This README is just the pitch.
Comparison
| CLAUDART | Mem0 | Zep | LangMem | Understand-Anything | MemPalace | |
|---|---|---|---|---|---|---|
| Setup | curl | bash | vector DB + Docker + OpenAI key | Neo4j + managed cloud | PostgreSQL + pgvector | curl | bash or plugin | pip install + 300 MB model |
| Human-readable | ✅ | ❌ | ❌ | ❌ | ⚠️ JSON + dashboard | ⚠️ verbatim text, binary DB |
| Works offline / air-gapped | ✅ | ❌ | ❌ | ❌ | ❌ LLM required | ✅ |
| PR-reviewable memory | ✅ | ❌ | ❌ | ❌ | ✅ JSON committed to git | ❌ ChromaDB + SQLite binary |
| Tool support | Claude Code, Codex CLI | API only | API only | LangGraph only | Claude, Codex, Cursor, Copilot, Gemini + 6 more | Claude Code, Codex CLI, Gemini CLI, MCP-compatible |
Markdown-in-the-repo won — AGENTS.md already appears in ~20k public repos. CLAUDART makes the convention structured and adds the workflow around it: orientation, planning, learning, hygiene, and review.
License
MIT. See LICENSE. Contributions welcome — see CONTRIBUTING.md.