Pure Markdown · Zero Dependencies · MIT Open Source

Your AI's memory,
governed.

AGENTS.md tells your AI what your project is.
MUSE tells it who it is, what it remembers, and what it can do.

View on GitHub Quick Install →
65Skills
10AI Tools
0Dependencies

Your AI forgets. Every. Single. Time.

Without MUSE

  • New conversation = re-explain your entire codebase
  • 3 hours in, context fills up, quality tanks
  • Yesterday's fix? Reintroduced today
  • Switch projects, AI mixes up everything
  • Forget to save? All progress gone

With MUSE

  • /resume — full context restored in 5 seconds
  • /bye — one-click save, zero input needed
  • /distill — lessons extracted to long-term memory
  • Role isolation — dev, QA, marketing stay separate
  • Auto-save every 15 rounds — crash-proof

🎯 Benchmark: /resume loads 16.9× more context than a bare .cursorrules cold start | 15/15 governance features vs 0/15 for format-spec-only setups

Run ./scripts/benchmark.sh on your own project to verify

What MUSE actually does

Six building blocks. All Markdown. All git-tracked.

Two-Layer Memory

Daily logs (memory/) capture what happened. MEMORIES.md distills what matters. Short-term meets long-term.

Role Isolation

Build, QA, Growth, Strategy — each stays in its own lane. Your AI won't suggest marketing while you're debugging.

Directive Queue

Async messaging across roles with delivery tracking. Strategy decides, Build ships, QA verifies — all through Markdown.

MCP Server

6 JSON-RPC tools in a single Bash script. Claude Code reads your project status through the standard protocol. No Node. No Python.

65 Reusable Skills

Git workflows, debugging, security guard, coordinator mode, iterative retrieval, deep-dive analysis — pre-built skills your AI loads on demand.

Crash-Proof Auto-Save

Silent checkpoints every 15 rounds. Context blowout? /resume crash picks up where you left off. At most 10 rounds lost.

Highlights

The features that make MUSE different

Self-Healing SOP

6 iterations and counting. Every governance bug found in production is patched back into /bye and /resume — role drift, memory loss, cross-project misrouting — all systematically fixed.

/sprint Pipeline

7-phase feature dev workflow: Think → Plan → Build → Review → Test → Ship → Reflect. Each phase loads the right skill automatically.

Git Security Guard

Pre-commit secret leak prevention. Catches API keys, tokens, private keys, .env files before they enter git history. Born from a real incident.

Online Dashboard

Visit muse.mythslabs.ai/dashboard — load your project, see roles, memory, skills visualized. 100% client-side.

Skill Manifest

One script indexes all 65 skills into manifest.json — name, category, size, assets. Machine-readable. CI-friendly --check mode catches stale indexes.

Graph Memory

UPDATES supersede old facts with ~~strikethrough~~. EXTENDS append sub-details. DERIVES link inferences. Auto-Expiry tags entries as [PERMANENT], [TEMPORAL], or [EPISODIC] — expired memories fade automatically.

Five seconds to full context

/resume → work → /bye — that's the whole workflow

MUSE Demo — /resume → boot → work → /ctx → /bye

Pick your tool. One command. Done.

Auto-converts to the right format for each tool

Claude Code
install.sh --tool claude
Cursor
install.sh --tool cursor
Windsurf
install.sh --tool windsurf
Gemini CLI
install.sh --tool gemini
Codex CLI
install.sh --tool codex
Copilot
install.sh --tool copilot
Aider
install.sh --tool aider
Antigravity
install.sh --tool antigravity
OpenClaw
install.sh --tool openclaw
OpenCode
install.sh --tool opencode

Up and running in 60 seconds

1

Clone MUSE

git clone https://github.com/myths-labs/muse.git
2

Install to your project

cd muse && ./scripts/install.sh --target /path/to/your-project
3

Start using

/resume → work → /bye

MUSE vs. format specs vs. memory infra

Format specs tell your AI the rules. Memory infra stores embeddings. MUSE governs the whole system.

Feature MUSE mem0 memsearch lossless-claw
Dependencies Zero Python + DB Milvus + Node SQLite + DAG
Format Pure Markdown Vector DB Vector DB Custom DAG
Tool Support 10 tools API only Claude Code OpenClaw
Role System ✅ Multi-role — — —
Cross-role Directives ✅ 📡 Protocol — — —
MCP Server ✅ Built-in — ✅ —
Graph Memory ✅ UPDATES/EXTENDS/DERIVES Vector similarity Vector similarity DAG nodes
Smart Forgetting ✅ Auto-Expiry tags TTL-based — —
Self-Healing SOP ✅ 6 iterations — — —
Digital Twin ✅ 5-dimension profile — — —
License MIT Apache 2.0 MIT MIT

The entire interface

/resume

Boot — auto-restore context & start work

/bye

One-click wrap-up — save, sync, archive

/sprint

Feature pipeline: Think→Plan→Build→Review→Test→Ship

/retro

Weekly retrospective from memory + git history

/ctx

Context health check (🟢🟡🔴)

/distill

Condense daily logs → long-term memory

/sync

Cross-role synchronization

/launch

Launch Day protocol — multi-platform social broadcast

/model

Switch AI model & update preferences

/release

Open-source release — version bump, tag, GitHub Release

/settings

Change language, model, preferences

"I'm a solo founder running 3 products with AI. 15-20 conversations a day. Without MUSE, I'd spend half my time re-explaining context."

JC Creator of MUSE · @SunshiningDay