MeOS is not a memory dump.
It is not a roleplay profile pack.
It is not just a pile of prompts.

It is a file-first system for making a person reusable to future agents.

Supports Codex, Claude Code, OpenCode, and OpenClaw.

If you are tired of repeating the same standards, the same taste, and the same corrections in every new agent session, MeOS is the layer that turns that repetition into reusable local assets.

✨ Why MeOS

Most systems in this space do one part well:

• distill a person into a prompt
• extract reusable skills from interaction logs
• keep reusable instructions around

MeOS focuses on the layer in between: the long-lived operating layer.

| Typical approach | What it usually misses | What MeOS does instead | |---|---|---| | Memory dump | Too much noise, weak reuse | Promotes only stable patterns into assets | | Persona prompt | Sounds similar, works shallowly | Stores workflows, standards, principles, and corrections | | One-shot profile summary | Becomes stale quickly | Supports init, refresh, and apply as an ongoing lifecycle | | Private local notebook | Hard to reuse across tools | Uses a shared SKILL.md shape that multiple runtimes can load |

The core idea is simple:

do not merely remember the owner, make the owner reusable

🧩 What You Actually Get

| Asset type | Examples | What it improves | |---|---|---| | 🛠 Work standards | coding rules, review bar, acceptance lines | technical quality and consistency | | 🧭 Workflow assets | debug order, architecture review sequence, delivery checklist | how the agent approaches the task | | 🧠 Thought style | reasoning patterns, trade-off habits, decision style | planning and judgment quality | | 🎨 Taste and preferences | output structure, UI taste, response style | how results look and read | | ✍️ Corrections | explicit overrides, things the owner rejected before | prevents repeated misalignment | | 📚 Knowledge assets | stable facts, domain understanding, reusable experience | task context that survives beyond one chat |

⚙️ How MeOS Works

| Mode | Purpose | Reads first | Writes back | |---|---|---|---| | 🧱 init | Build the first asset set from approved local material | source policy, extraction SOP, promotion policy, privacy policy | initial assets plus evidence | | 🔁 refresh | Update existing assets with new material | extraction SOP, promotion policy, correction policy | merged updates, conflicts, corrections | | 🎯 apply | Use existing assets during a live task | only the minimum relevant assets | only stable new information |

apply is the most important mode. That is where MeOS stops being an archive and starts becoming useful.

🚀 Quick Start

Fastest path: install from npm

npm install -g @researai/meos
meos install --runtime codex
meos doctor

Then use it immediately:

Use meos in apply mode for this task. Read only the minimum relevant assets and use them to shape reasoning, workflow, and output.

Repository path: clone and install locally

git clone https://github.com/ResearAI/MeOS.git
cd MeOS
bash install.sh --runtime codex
python3 installer.py doctor

📦 Install And Use

Install from npm

This is the recommended path if you just want to use MeOS:

npm install -g @researai/meos

Core commands:

| Command | What it does | |---|---| | meos install --runtime codex | install MeOS into Codex | | meos install --runtime claude | install MeOS into Claude Code | | meos install --runtime openclaw --force | install MeOS into OpenClaw | | meos install --runtime opencode | install MeOS into OpenCode | | meos doctor | check repository and target install paths | | meos print-prompts --lang en | print suggested init / refresh / apply prompts | | meos print-prompts --lang zh | print the same prompts in Chinese |

Install from the repository

Use this if you want to work on MeOS itself:

The installer does not publish the whole repository. It installs the contents of ./SKILL/ into the runtime skill directory meos/.

--mode auto is the default. It uses a runtime-safe install strategy:

• packaged installs from npm default to copy, so mutable local assets do not live inside node_modules
• local repository installs default to copy for OpenClaw and symlink for the other runtimes

| Runtime | Recommended command | Note | |---|---|---| | Codex | bash install.sh --runtime codex | simplest path for local skill use | | Claude Code | bash install.sh --runtime claude | use lowercase meos skill dir | | OpenClaw | bash install.sh --runtime openclaw --force | prefers copied skill directories | | OpenCode | bash install.sh --runtime opencode | install into one compatible path only |

You can also use the npm wrapper:

npm install -g @researai/meos
meos install --runtime codex

For local development of this repository:

npm install -g .

Common install targets

| Runtime | Recommended command | Note | |---|---|---| | Codex | meos install --runtime codex | simplest path for local skill use | | Claude Code | meos install --runtime claude | use lowercase meos skill dir | | OpenClaw | meos install --runtime openclaw --force | prefers copied skill directories | | OpenCode | meos install --runtime opencode | install into one compatible path only |

If you are replacing an older install instead of doing a first install, add --force.

Use MeOS in a real task

The three core usage modes are:

| Mode | When to use it | Prompt | |---|---|---| | init | build the first asset set | Use meos in init mode. Build the first sanitized operating-layer assets from the available local source material. | | refresh | update existing assets | Use meos in refresh mode. Refresh the existing MeOS assets from new local material and only promote stable rules. | | apply | use existing assets during work | Use meos in apply mode for this task. Read only the minimum relevant assets and write back only stable new information. |

Runtime-specific usage examples

Codex:

Use meos in apply mode for this task. Read only the minimum relevant assets and use them to shape reasoning, workflow, and output.

Claude Code:

/meos
Apply MeOS for this task. Read only the minimum relevant assets and use them to shape reasoning, workflow, and output.

🖥 Runtime Setup

Codex

Codex supports skill directories such as .agents/skills/ and ~/.agents/skills/.

Manual install:

mkdir -p ~/.agents/skills
ln -s /path/to/MeOS/SKILL ~/.agents/skills/meos

Codex can trigger MeOS explicitly by name or implicitly through the skill description.

Claude Code

Claude Code supports ~/.claude/skills/<skill-name>/SKILL.md and .claude/skills/<skill-name>/SKILL.md.

Manual install:

mkdir -p ~/.claude/skills
ln -s /path/to/MeOS/SKILL ~/.claude/skills/meos

Typical use:

/meos
Apply MeOS for this task. Read only the minimum relevant assets and use them to shape reasoning, workflow, and output.

Claude Code + MiniMax

If you want Claude Code to use the MiniMax Anthropic-compatible endpoint, keep a local ~/.claude/settings.json like this:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.minimaxi.com/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "${MINIMAX_API_KEY}",
    "API_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

Verify:

claude -p --model MiniMax-M2.7 'Respond with exactly CLAUDE_MINIMAX_OK.'

OpenClaw

OpenClaw supports ~/.openclaw/skills, ~/.agents/skills, <workspace>/.agents/skills, and <workspace>/skills.

Important behavior from manual testing:

• OpenClaw skips symlinked skill roots whose resolved realpath escapes the configured root.
• In practice, external symlink installs are unreliable.
• The safest install is a copied directory under <workspace>/skills/meos or ~/.openclaw/skills/meos.

Recommended manual install:

mkdir -p <workspace>/skills
cp -a /path/to/MeOS/SKILL <workspace>/skills/meos

Verify:

openclaw skills info meos
openclaw skills list | rg meos

OpenCode

OpenCode searches several compatible skill locations:

• .opencode/skills/<name>/SKILL.md
• ~/.config/opencode/skills/<name>/SKILL.md
• .claude/skills/<name>/SKILL.md
• ~/.claude/skills/<name>/SKILL.md
• .agents/skills/<name>/SKILL.md
• ~/.agents/skills/<name>/SKILL.md

Choose one install path only:

mkdir -p ~/.config/opencode/skills
ln -s /path/to/MeOS/SKILL ~/.config/opencode/skills/meos

If your provider or proxy does not support OpenCode's default secondary title model, set small_model explicitly:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "openai/gpt-5.4",
  "small_model": "openai/gpt-5.4"
}

Verify:

opencode run --model openai/gpt-5.4 --format json \
  'Use meos in apply mode for this task. Reply with exactly OPENCODE_SKILL_OK.'

🗂 What The Agent Reads In apply Mode

Paths below are relative to the installed skill root. Inside this repository, they live under SKILL/.

| Task type | Read first | Outcome | |---|---|---| | 🛠 Technical implementation | assets/live/work/, assets/live/thought-style/, assets/live/workflow/, assets/live/principles/ | follows your technical standards and execution order | | 🎨 UI / product work | assets/live/taste/, assets/live/work/, assets/live/workflow/, assets/live/corrections/ | preserves your taste and presentation bar | | 🔬 Research / writing | assets/live/work/, assets/live/thought-style/, assets/live/principles/, assets/live/knowledge/, assets/live/workflow/ | uses your structure, reasoning, and domain framing | | 💬 Style-sensitive replies | assets/live/preferences/, assets/live/corrections/ | matches preferred response shape and wording |

If assets/live/corrections/ conflicts with another layer, correction wins.

🔒 Promotion And Privacy Rules

Promotion flow

1. collect local material
2. classify the source
3. extract high-signal candidate facts
4. keep uncertain items in evidence/
5. promote only stable items into assets/live/

Promote when one of these is true:

• explicit user statement
• repeated pattern across contexts
• explicit correction or reinforcement

Keep the item in evidence/ when it is:

• one-off
• noisy
• too context-specific
• too sensitive

Maintenance lifecycle

MeOS should evolve through:

• add
• merge
• downgrade
• discard

The point is to keep assets clean and maintainable, not to accumulate prompt clutter.

Privacy boundary

| Safe to publish | Keep local by default | |---|---| | README.md | SKILL/private/ | | README_ZH.md | SKILL/evidence/ | | assets/branding/ | SKILL/runtime/ | | assets/readme/ | SKILL/assets/live/ | | SKILL/references/ | raw imported material | | SKILL/schemas/ | secrets and tokens | | SKILL/assets/templates/ | workstation-specific notes | | SKILL/assets/examples/ | private raw transcripts |

Do not commit:

• tokens
• API keys
• personal identifiers
• raw connector ids
• unnecessary private paths
• raw private transcripts

🏗 Repository Layout

MeOS/
├── README.md
├── README_ZH.md
├── LICENSE
├── assets/
│   ├── branding/
│   └── readme/
├── SKILL/
│   ├── SKILL.md
│   ├── references/
│   ├── schemas/
│   ├── assets/
│   │   ├── templates/
│   │   ├── examples/
│   │   └── live/
│   ├── evidence/
│   ├── runtime/
│   └── private/
├── install.sh
├── installer.py
├── package.json
└── bin/

The important pattern is:

• public project materials stay at the repository root
• everything the runtime skill needs lives under SKILL/
• the installer publishes SKILL/ into runtime skill directories
• local-only owner calibration stays in SKILL/assets/live/, SKILL/evidence/, SKILL/private/, and SKILL/runtime/

📌 Key Skill References

The highest-value files inside SKILL/ are:

• SKILL/SKILL.md
• SKILL/references/source-locations.md
• SKILL/references/extraction-sop.md
• SKILL/references/promotion-policy.md
• SKILL/references/privacy-policy.md
• SKILL/references/writeback-policy.md
• SKILL/references/apply-task-map.md

🛤 Current Direction

MeOS already has:

• a cross-tool SKILL/ package layout
• extraction, promotion, privacy, and writeback references
• JSON schemas for durable entries
• example and template asset trees
• installer paths for Codex, Claude Code, OpenCode, and OpenClaw

The next step is not "add more prompt text". It is to keep improving the assets, examples, and public presentation without leaking private local history.

📚 Citation

If MeOS is used in work related to personal alignment, style distillation, or operating-layer asset maintenance, you may also want to cite:

@inproceedings{
zhu2025personality,
title={Personality Alignment of Large Language Models},
author={Minjun Zhu and Yixuan Weng and Linyi Yang and Yue Zhang},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=0DZEs8NpUH}
}