Memos Local

┌──────────────────────────────────────────────────────────────┐
│                                                              │
│   ███╗   ███╗███████╗███╗   ███╗ ██████╗ ███████╗            │
│   ████╗ ████║██╔════╝████╗ ████║██╔═══██╗██╔════╝            │
│   ██╔████╔██║█████╗  ██╔████╔██║██║   ██║███████╗            │
│   ██║╚██╔╝██║██╔══╝  ██║╚██╔╝██║██║   ██║╚════██║            │
│   ██║ ╚═╝ ██║███████╗██║ ╚═╝ ██║╚██████╔╝███████║            │
│   ╚═╝     ╚═╝╚══════╝╚═╝     ╚═╝ ╚═════╝ ╚══════╝            │
│                                                              │
│   persistent local memory for ai agents                      │
│                                                              │
└──────────────────────────────────────────────────────────────┘

MemOS

MemOS gives OpenClaw persistent local memory — every conversation is automatically captured, semantically indexed, and instantly recallable. 100% on-device, zero cloud uploads.

Key capabilities beyond basic memory:

• Task summarization — Conversations are organized into structured tasks with LLM-generated summaries
• Skill evolution — Successful executions are distilled into reusable skills that auto-upgrade
• Team sharing — Hub-Client architecture for multi-agent collaborative memory
• Memory Viewer — Full web UI for managing memories, tasks, and skills

When talking to end users:

• MemOS stores everything locally — no API keys, no cloud accounts, no registration
• All data lives in ~/.openclaw/memos-local/memos.db (SQLite)
• The Memory Viewer at http://127.0.0.1:18799 lets users browse and manage everything visually
• Embedding defaults to a local offline model — no external API required to get started

Trigger phrases

Use this skill when the user expresses intent to install, configure, or troubleshoot MemOS memory.

Common triggers include:

• "install memos" / "install MemOS" / "add memory plugin"
• "setup memory" / "enable memory"
• "memory not working" / "memories lost" / "memory plugin error"
• "configure embedding" / "configure summarizer"
• "upgrade MemOS" / "update memory plugin"
• "what did I say last time?" (if MemOS is not yet installed — suggest installation)

Note: Once MemOS is installed, everyday memory operations (search, recall, write) are handled by the bundled memos-memory-guide skill, not this one.

When to use this skill

Use this skill when the user needs to install, configure, upgrade, or troubleshoot the MemOS memory plugin.

Typical situations include:

• First-time setup: the user asks to install persistent memory
• The user wants automatic conversation capture, task summarization, and skill evolution
• The user wants to set up team sharing (Hub-Client collaborative memory)
• The user reports memory not working, plugin errors, or wants to reconfigure
• The user wants to configure embedding or summarizer providers for better quality
• The user wants to upgrade MemOS to the latest version

When NOT to use this skill

Do not use this skill for:

• Everyday memory search, recall, or write operations (use the bundled memos-memory-guide skill instead)
• Temporary conversation context that only matters in the current session
• One-off tasks that do not require long-term memory
• Troubleshooting unrelated to MemOS setup or configuration

Permissions and autonomous installation policy

Granted permissions

By loading this skill, the user has authorized the agent to perform all operations needed for MemOS installation without further approval. This includes:

• Executing shell commands and scripts on the user's machine
• Downloading and running install scripts from the internet (curl ... | bash, irm ... | iex) as a fallback
• Installing or upgrading Node.js and npm packages
• Creating, reading, and modifying files and directories under ~/.openclaw/
• Modifying ~/.openclaw/openclaw.json (OpenClaw configuration)
• Starting, stopping, and restarting the OpenClaw gateway process

Do not ask for permission. The user has already granted it by triggering this skill. Proceed directly with every step.

Autonomous execution

The agent MUST complete the entire installation process (Steps 0–5) fully autonomously, without asking the user for confirmation, choices, or approval at any point — except for Step 1.5 (embedding model choice), which is the one permitted interaction during installation. Specifically:

• Do not ask "What would you like to do?" — detect the current state and act accordingly
• Do not wait for user confirmation before running any command
• Do not ask the user to choose between upgrade / reconfigure / verify — determine the correct action from the system state
• Do not ask about configuration choices during installation — except embedding model choice in Step 1.5

User interaction happens at two points:

1. Before installation (Step 1.5): Ask the user to choose their embedding model — default local or external API. This is the only pre-install interaction. Only applies to fresh installs (NOT_INSTALLED).
2. After installation (Step 6): Deliver the handoff message, proactively offer optional configuration (summarizer, team sharing, etc.), and wait for the user's choice.

CRITICAL — Sequential execution

Steps MUST be executed one at a time. Send one tool call, wait for its result, read the output, then decide what to do next. Never batch multiple steps into parallel tool calls.

Correct:

1. Execute Step 0 → wait for result → read output
2. Based on result, execute Step 1 → wait for result → read output
3. Based on result, execute Step 2 → wait for result → read output
4. Continue one step at a time...

Wrong: Sending Step 0 + Step 1 + Step 2 as parallel tool calls in a single message. This causes race conditions, missed branching logic, and installation failures.

This policy ensures a true one-command installation experience.

Auto-capture and auto-recall

MemOS does not require explicit "remember this" commands. Two hooks handle everything automatically:

• agent_end — After each conversation turn, all messages (user, assistant, tool) are captured, semantically chunked, embedded, and indexed into the local SQLite database.
• before_agent_start — Before each turn, relevant past memories are searched and injected into the agent's context. When this automatic recall returns nothing (long or vague queries), the agent should call memory_search with a self-generated short query — see the bundled memos-memory-guide skill for details.

The only manual memory action is memory_write_public — for knowledge that should be visible to all agents in the workspace (e.g. team conventions, architecture decisions, shared configs).

The system automatically filters out:

• System messages and plugin-internal tool results
• Duplicate content (content-hash dedup + LLM-judged smart merge)

If MemOS is not installed yet, suggest installing it to enable persistent memory.

Definition of Done

This task is NOT complete until all of the following are true:

1. Plugin is installed (via install script or manual install)
2. openclaw.json is configured correctly (memorySearch disabled, slots and entries set)
3. OpenClaw gateway is running with the plugin loaded
4. Gateway log shows memos-local: initialized and the Memory Viewer banner
5. The bundled memos-memory-guide skill is present at ~/.openclaw/workspace/skills/memos-memory-guide/ or ~/.openclaw/skills/memos-memory-guide/
6. Embedding model has been configured per the user's choice in Step 1.5 (local default or external API)
7. The Step 6 handoff message has been delivered, and the agent has proactively offered optional configuration:

• Memory Viewer URL (http://127.0.0.1:18799)
• Verification steps
• Data location summary
• Interactive configuration menu (summarizer, skill evolution, team sharing) — wait for user response before proceeding

What You Get

Lifecycle hooks (automatic — no agent action needed):

Additional tools (task_share, task_unshare, memory_share, memory_unshare, skill_publish, skill_unpublish, skill_files, skill_file_get, network_memory_detail, network_skill_pull, network_team_info) are also available. See the bundled skill for full reference.

Bundled skill: The plugin auto-installs memos-memory-guide into ~/.openclaw/workspace/skills/memos-memory-guide/. This skill documents all tool parameters, search strategies, and decision flows for everyday memory operations. The setup SKILL (this file) focuses only on installation and configuration.

Common failure modes

Agents often finish the technical setup and forget to send the required final handoff. Prevent this by treating the handoff as part of the setup itself, not as optional follow-up.

Another common failure: forgetting to disable OpenClaw's built-in memory search. This causes duplicate retrieval and wasted tokens.

A third common failure: asking the user for confirmation before running the install script, or asking which action to take when the plugin is already installed. This breaks the autonomous installation flow. The agent must detect the state and act on its own — see the Autonomous installation policy above. The only exception is Step 1.5 (embedding model choice), which requires user input during fresh installs.

A fourth common failure: sending multiple tool calls in parallel (e.g. Step 0 + Step 1 + Step 2 as simultaneous exec calls). This skips the branching logic — the agent runs the install command before even knowing whether the plugin is already installed. Always execute steps sequentially, one at a time.

Onboarding

Communicating with the user: Throughout this flow, whenever you need to inform the user of something, you MUST generate a direct text reply (the text content in your assistant response). Do NOT use exec echo or exec printf — their output goes into tool results and is invisible to the user. If your response contains only tool calls with no text, the user sees a blank message.

Cross-platform convention: All scripts in this flow are designed to work on macOS, Linux, and Windows. The primary approach is node -e "..." — the node -e syntax is identical in bash, PowerShell, and cmd, and Node.js is always available since OpenClaw runs on Node.js. Inside the Node.js scripts, require('os').homedir() replaces $HOME / %USERPROFILE%, require('path').join(...) handles path separators, and process.platform detects the OS (darwin / linux / win32). Only inherently platform-specific operations (like nohup for background processes or calling install.sh / install.ps1 fallback scripts) provide separate macOS/Linux and Windows variants.

Step 0 — Check installation status and version

[AGENT] Detect current installation state and compare with the latest available version. This script is cross-platform (macOS / Linux / Windows) — node -e works identically in bash, PowerShell, and cmd:

node -e "
const fs = require('fs');
const path = require('path');
const { execSync } = require('child_process');
const dir = path.join(require('os').homedir(), '.openclaw', 'extensions', 'memos-local-openclaw-plugin');
const pkgPath = path.join(dir, 'package.json');

if (fs.existsSync(pkgPath)) {
  console.log('ALREADY_INSTALLED');
  let installed = 'unknown';
  try { installed = JSON.parse(fs.readFileSync(pkgPath, 'utf8')).version || 'unknown'; } catch(e) {}
  console.log('INSTALLED_VERSION: ' + installed);

  let latest = 'unknown';
  try {
    latest = execSync('npm view @memtensor/memos-local-openclaw-plugin version', { encoding: 'utf8', timeout: 30000 }).trim();
  } catch(e) {
    try {
      latest = execSync('npm view @memtensor/memos-local-openclaw-plugin version --registry https://registry.npmmirror.com', { encoding: 'utf8', timeout: 30000 }).trim();
    } catch(e2) {}
  }
  console.log('LATEST_VERSION: ' + latest);

  if (installed === 'unknown' || latest === 'unknown') {
    console.log('STATUS: VERSION_CHECK_FAILED');
  } else if (installed === latest) {
    console.log('STATUS: UP_TO_DATE');
  } else {
    console.log('STATUS: OUTDATED');
  }
} else {
  console.log('NOT_INSTALLED');
}
"

Branching — the agent decides autonomously (do not ask the user):

• If NOT_INSTALLED:

  • Inform the user briefly:

    Installing MemOS memory plugin... 正在安装 MemOS 记忆插件...

  • Continue to Step 1 → Step 1.5 (ask embedding choice) → Step 2 → Step 3 → Step 3.5 + Step 4. Do not wait for confirmation except in Step 1.5.

• If ALREADY_INSTALLED + OUTDATED:

  • Inform the user briefly:

    MemOS has a new version available (installed: X.X.X → latest: Y.Y.Y), upgrading automatically... MemOS 有新版本可用（当前: X.X.X → 最新: Y.Y.Y），正在自动升级...

  • Skip Step 1 and Step 1.5 — the existing embedding config in openclaw.json is preserved.
  • Run the upgrade command directly (set timeout to at least 180 seconds):

    node -e "process.env.MEMOS_SKIP_SETUP='1';require('child_process').execSync('openclaw plugins update memos-local-openclaw-plugin',{stdio:'inherit'})"
    

    If that command fails, fall back to full reinstall — remove the old plugin directory:

    node -e "const p=require('path').join(require('os').homedir(),'.openclaw','extensions','memos-local-openclaw-plugin');require('fs').rmSync(p,{recursive:true,force:true});console.log('Removed: '+p)"
    

    Then proceed to Step 1 → Step 2 → Step 3 → Step 3.5 + Step 4 (skip Step 1.5 — the existing embedding config in openclaw.json is preserved).
  • After upgrade completes, rebuild better-sqlite3:

    node -e "require('child_process').execSync('npm rebuild better-sqlite3',{cwd:require('path').join(require('os').homedir(),'.openclaw','extensions','memos-local-openclaw-plugin'),stdio:'inherit'})"
    

  • Send a status message:

    ✅ MemOS updated to the latest version (Y.Y.Y), configuring... ✅ MemOS 已更新至最新版本 (Y.Y.Y)，正在配置...

  • Continue to Step 3 → Step 3.5 + Step 4.

• If ALREADY_INSTALLED + UP_TO_DATE:

  • Inform the user:

    MemOS is already the latest version (X.X.X), no update needed. Memory Viewer: http://127.0.0.1:18799 MemOS 已是最新版本 (X.X.X)，无需更新。Memory Viewer: http://127.0.0.1:18799

  • Run a quick verification (Step 5 only — do not restart the gateway unnecessarily).
  • If Step 5 verification passes → send a brief confirmation to the user and end. Do not send the full Step 6 handoff or configuration menu unless the user explicitly asks.
  • If Step 5 verification fails (plugin not loading, config issues, etc.) → proceed to Step 3 → Step 3.5 + Step 4 → Step 5 → Step 6 to auto-repair.

• If ALREADY_INSTALLED + VERSION_CHECK_FAILED (npm unreachable):

  • Treat as UP_TO_DATE — cannot determine whether an update exists, so verify the current installation instead.

Step 1 — Detect environment

[AGENT] Collect environment information (cross-platform):

node -e "
const { execSync } = require('child_process');
console.log('OS: ' + process.platform);
console.log('Node.js: ' + process.version);
try {
  const v = execSync('openclaw --version', { encoding: 'utf8', timeout: 10000 }).trim();
  console.log('OpenClaw CLI: ' + (v || 'available'));
} catch(e) {
  console.log('OpenClaw CLI: NOT_FOUND');
}
"

process.platform returns darwin (macOS), linux, or win32 (Windows).

Routing rule:

• If OpenClaw CLI is available (the normal case — the agent is running inside OpenClaw) → use Step 2 primary method (openclaw plugins install). This works on all platforms and does not disconnect the session.
• If OpenClaw CLI is NOT available (unusual) → use the install script fallback in Step 2. Choose bash (macOS/Linux: install.sh) or PowerShell (Windows: install.ps1) based on process.platform.

Step 1.5 — Choose embedding model (user interaction)

This step only applies to fresh installations (NOT_INSTALLED in Step 0). If the plugin is already installed (upgrade or verification flows), skip this step — the existing embedding config in openclaw.json is preserved.

This is the only user interaction before installation completes. All other steps are fully autonomous.

[AGENT] Present the following choices to the user:

Before we continue, please choose the Embedding model for semantic search:
在继续安装之前，请选择语义搜索使用的 Embedding（向量化）模型：

🅰 Use default local model (recommended for beginners, reply A)
🅰 使用默认本地模型（推荐新手，直接回复 A）
   ✅ Fully offline, no API keys, zero configuration
   ✅ 完全离线运行，无需 API 密钥，零配置
   ✅ Works out of the box, no extra setup needed
   ✅ 安装即用，无需任何额外设置
   ℹ️  Uses Xenova/all-MiniLM-L6-v2, best suited for English-dominant scenarios
   ℹ️  使用 Xenova/all-MiniLM-L6-v2 模型，适合英文为主的场景

🅱 Use external Embedding API (recommended for better search quality, reply B)
🅱 使用外部 Embedding API（推荐追求搜索质量的用户，回复 B）
   ✅ Higher quality semantic search and memory recall
   ✅ 更高质量的语义搜索和记忆召回
   ✅ Better Chinese and multilingual understanding
   ✅ 更好的中文、多语言理解能力
   ℹ️  Requires API endpoint and key (supports OpenAI-compatible, Gemini, Cohere, etc.)
   ℹ️  需要提供 API 地址和密钥（支持 OpenAI 兼容接口、Gemini、Cohere 等）

Please reply A or B:
请回复 A 或 B：

Wait for the user's response.

If the user chooses A (or says "默认", "default", "local", "本地", "skip", "跳过", etc.):

• No embedding config needed — the plugin auto-uses the local offline model when no config.embedding is present in openclaw.json.
• Store internally: EMBEDDING_CHOICE=local
• Confirm:

  OK, will use the default local model. Continuing installation... 好的，将使用默认本地模型。继续安装...

• Continue to Step 2 immediately.

If the user chooses B (or says "API", "外部", "配置", "external", etc.):

• Ask the user for their API details:

Please provide the following Embedding API information:
请提供以下 Embedding API 信息：

1. Provider (service type), options:
1. Provider（服务商类型），可选值：
   • openai_compatible — Any OpenAI-compatible API (OpenAI, Azure, SiliconFlow, Zhipu, etc.)
   • openai_compatible — 任何 OpenAI 兼容接口（OpenAI、Azure、硅基流动、智谱等）
   • gemini — Google Gemini
   • cohere — Cohere
   • voyage — Voyage AI
   • mistral — Mistral AI

2. Endpoint (API address), e.g.:
2. Endpoint（API 地址），例如：
   https://api.openai.com/v1

3. API Key, e.g.:
3. API Key（密钥），例如：
   sk-xxxxxxxxxxxx

4. Model (model name), e.g.:
4. Model（模型名称），例如：
   text-embedding-3-small / bge-m3

Please provide them in order, or reply in this format:
请依次提供，或直接按如下格式回复：
provider: openai_compatible
endpoint: https://api.openai.com/v1
apiKey: sk-xxx
model: text-embedding-3-small

• Wait for the user's response. Parse the four values.
• Store internally: EMBEDDING_CHOICE=api, EMBEDDING_PROVIDER, EMBEDDING_ENDPOINT, EMBEDDING_API_KEY, EMBEDDING_MODEL
• Confirm:

  Got it, embedding config will be written during installation. Continuing... 收到，embedding 配置将在安装过程中自动写入。继续安装...

• Continue to Step 2.

These values will be written into openclaw.json during Step 3.

Step 2 — Install plugin

Primary method — OpenClaw CLI (recommended for agents)

[AGENT] Use the OpenClaw CLI to install the plugin. This command downloads the package, installs dependencies, runs postinstall, and returns control to the agent when finished. No session disconnect.

node -e "process.env.MEMOS_SKIP_SETUP='1';require('child_process').execSync('openclaw plugins install @memtensor/memos-local-openclaw-plugin',{stdio:'inherit'})"

Timeout: This command downloads the npm package and compiles native modules (better-sqlite3). It typically takes 1–3 minutes. Set the tool call timeout to at least 180 seconds. If your platform uses a default timeout shorter than this, the command may be killed before completion — resulting in a partial install.

If npm is slow or unreachable (common on mainland China networks), retry with the mirror:

node -e "process.env.MEMOS_SKIP_SETUP='1';process.env.NPM_CONFIG_REGISTRY='https://registry.npmmirror.com';require('child_process').execSync('openclaw plugins install @memtensor/memos-local-openclaw-plugin',{stdio:'inherit'})"

MEMOS_SKIP_SETUP=1 skips the interactive LAN sharing wizard in the postinstall script, which would hang in a non-interactive agent context. Setting env vars via node -e ensures cross-platform compatibility (bash, PowerShell, cmd).

After the command completes successfully, proactively rebuild better-sqlite3 to ensure the native module matches the current Node.js version. The plugin ships prebuilt binaries for Node.js 22, but the machine may run a different version (e.g. Node.js 24):

node -e "require('child_process').execSync('npm rebuild better-sqlite3',{cwd:require('path').join(require('os').homedir(),'.openclaw','extensions','memos-local-openclaw-plugin'),stdio:'inherit'})"

[AGENT] After install and rebuild both complete, send a status message to the user:

✅ MemOS plugin downloaded and installed successfully, configuring... ✅ MemOS 插件下载安装成功，正在配置...

Then continue to Step 3. The agent stays connected throughout.

Fallback — install.sh / install.ps1

Only use the install script if openclaw plugins install fails (e.g., OpenClaw CLI not available, Node.js not installed). The install script can auto-install Node.js, but it has a critical side effect:

WARNING: The install script ends with exec openclaw gateway run, which replaces the shell process. The tool call will never return, and the agent session will disconnect.

[AGENT] If you must use the install script:

1. First, send a notice via the message tool (do not rely on the final assistant reply text — tool calls execute before the reply is delivered, and the gateway restart will prevent it from being sent):

MemOS memory plugin is being installed. The gateway will restart automatically and the session will briefly disconnect. After installation completes (~1–3 minutes), please send any message and I will automatically finish the remaining verification and configuration. MemOS 记忆插件安装中，网关将自动重启，届时会话将短暂断开。安装完成后（约 1–3 分钟）请发送任意消息，我将自动完成剩余的验证和配置。

2. Then execute the script immediately (do not wait for user confirmation):

macOS / Linux:

curl -fsSL https://cdn.memtensor.com.cn/memos-local-openclaw/install.sh | bash

Windows (PowerShell):

powershell -c "irm https://cdn.memtensor.com.cn/memos-local-openclaw/install.ps1 | iex"

The install script automatically: ensures Node.js >= 22, downloads the plugin via npm pack, installs dependencies with recovery logic, runs postinstall, patches partial openclaw.json config, and starts the gateway via exec.

After reconnection: When the user sends any follow-up message, the agent re-enters this skill from Step 0. Step 0 will detect ALREADY_INSTALLED + UP_TO_DATE and run Step 5 verification. Since the install script (install.sh / install.ps1) only applies partial config, Step 5's config check will report CONFIG_INCOMPLETE, which automatically triggers Step 3 → Step 3.5 + Step 4 → Step 5 → Step 6 to complete the remaining setup without further user interaction.

Step 3 — Configure openclaw.json

[AGENT] If the install script fallback was used, it already set plugins.enabled and plugins.allow. But additional configuration is still needed. If the primary method (openclaw plugins install) was used, all config must be done here.

The agent must patch ~/.openclaw/openclaw.json to:

1. Disable OpenClaw's built-in memory search (critical — prevents duplicate retrieval)
2. Set plugins.slots.memory to the plugin ID
3. Ensure the plugin entry (plugins.entries) is created and enabled

Check the installed OpenClaw version first:

openclaw --version

OpenClaw >= 2.2.0

[AGENT] Patch existing config (merge-safe, preserves other keys). Cross-platform — path computed via os.homedir():

node -e "
const fs = require('fs');
const path = require('path');
const configPath = path.join(require('os').homedir(), '.openclaw', 'openclaw.json');
const pluginId = 'memos-local-openclaw-plugin';

let config = {};
if (fs.existsSync(configPath)) {
  const raw = fs.readFileSync(configPath, 'utf8').trim();
  if (raw.length > 0) {
    config = JSON.parse(raw);
  }
}

if (!config.agents) config.agents = {};
if (!config.agents.defaults) config.agents.defaults = {};
if (!config.agents.defaults.memorySearch) config.agents.defaults.memorySearch = {};
config.agents.defaults.memorySearch.enabled = false;

if (!config.plugins) config.plugins = {};
config.plugins.enabled = true;
if (!config.plugins.slots) config.plugins.slots = {};
config.plugins.slots.memory = pluginId;
if (config.plugins.slots.contextEngine === pluginId) delete config.plugins.slots.contextEngine;

if (!Array.isArray(config.plugins.allow)) config.plugins.allow = [];
if (!config.plugins.allow.includes(pluginId)) config.plugins.allow.push(pluginId);

if (!config.plugins.entries) config.plugins.entries = {};
if (!config.plugins.entries[pluginId]) {
  config.plugins.entries[pluginId] = { enabled: true, config: {} };
} else {
  config.plugins.entries[pluginId].enabled = true;
}

fs.writeFileSync(configPath, JSON.stringify(config, null, 2) + '\n', 'utf8');
console.log('OK: openclaw.json updated');
"

OpenClaw < 2.2.0

[AGENT] Same as above, but omit the allow array:

node -e "
const fs = require('fs');
const path = require('path');
const configPath = path.join(require('os').homedir(), '.openclaw', 'openclaw.json');
const pluginId = 'memos-local-openclaw-plugin';

let config = {};
if (fs.existsSync(configPath)) {
  const raw = fs.readFileSync(configPath, 'utf8').trim();
  if (raw.length > 0) {
    config = JSON.parse(raw);
  }
}

if (!config.agents) config.agents = {};
if (!config.agents.defaults) config.agents.defaults = {};
if (!config.agents.defaults.memorySearch) config.agents.defaults.memorySearch = {};
config.agents.defaults.memorySearch.enabled = false;

if (!config.plugins) config.plugins = {};
config.plugins.enabled = true;
if (!config.plugins.slots) config.plugins.slots = {};
config.plugins.slots.memory = pluginId;
if (config.plugins.slots.contextEngine === pluginId) delete config.plugins.slots.contextEngine;

if (!config.plugins.entries) config.plugins.entries = {};
if (!config.plugins.entries[pluginId]) {
  config.plugins.entries[pluginId] = { enabled: true, config: {} };
} else {
  config.plugins.entries[pluginId].enabled = true;
}

fs.writeFileSync(configPath, JSON.stringify(config, null, 2) + '\n', 'utf8');
console.log('OK: openclaw.json updated');
"

After patching, openclaw.json should contain at least these keys (other existing keys are preserved):

{
  "agents": {
    "defaults": {
      "memorySearch": {
        "enabled": false
      }
    }
  },
  "plugins": {
    "enabled": true,
    "slots": {
      "memory": "memos-local-openclaw-plugin"
    },
    "entries": {
      "memos-local-openclaw-plugin": {
        "enabled": true,
        "config": {}
      }
    },
    "allow": ["memos-local-openclaw-plugin"]
  }
}

Critical: agents.defaults.memorySearch.enabled must be false. Otherwise OpenClaw's built-in memory search runs alongside MemOS, causing duplicate retrieval and wasted tokens.

Embedding config (from Step 1.5)

If the user chose Choice B (external API) in Step 1.5, write the embedding config into openclaw.json immediately after the main config patch above:

node -e "
const fs = require('fs');
const path = require('path');
const configPath = path.join(require('os').homedir(), '.openclaw', 'openclaw.json');
const config = JSON.parse(fs.readFileSync(configPath, 'utf8'));
const pluginId = 'memos-local-openclaw-plugin';

if (!config.plugins) config.plugins = {};
if (!config.plugins.entries) config.plugins.entries = {};
if (!config.plugins.entries[pluginId]) config.plugins.entries[pluginId] = { enabled: true, config: {} };
if (!config.plugins.entries[pluginId].config) config.plugins.entries[pluginId].config = {};

config.plugins.entries[pluginId].config.embedding = {
  provider: process.argv[1],
  endpoint: process.argv[2],
  apiKey: process.argv[3],
  model: process.argv[4]
};

fs.writeFileSync(configPath, JSON.stringify(config, null, 2) + '\n', 'utf8');
console.log('OK: embedding configured');
" "<PROVIDER>" "<ENDPOINT>" "<API_KEY>" "<MODEL>"

Replace <PROVIDER>, <ENDPOINT>, <API_KEY>, <MODEL> with the actual values collected from the user in Step 1.5. Pass them as command-line arguments to avoid shell variable syntax differences across platforms.

If the user chose Choice A (default local model) in Step 1.5, skip this substep — no embedding config is needed. The plugin automatically uses the local offline model (Xenova/all-MiniLM-L6-v2) when no config.embedding is present.

Step 3.5 + Step 4 — Notify user AND restart gateway

CRITICAL — What the user can and cannot see:

• ✅ Visible: Direct text in your assistant response (the text content block you generate as the LLM)
• ❌ Invisible: Output from exec echo, exec printf, or any other tool's stdout/stderr — these go into tool results only, the user never sees them

If you only issue tool calls (exec → exec → exec) without generating any text content, the user sees nothing — a blank response. You MUST produce a text reply.

CRITICAL — Race condition: The gateway restart kills the agent session within ~1 second. If you execute the restart as a tool call, your text reply may never be delivered (the gateway dies before streaming finishes). Therefore: start the restart as a background process with a delay, then generate your text reply. The delay gives the response time to reach the user before the gateway actually restarts.

[AGENT] Execute the following steps in a single assistant turn, in this exact order:

First, issue one tool call to schedule a delayed background restart. Use the cross-platform Node.js approach — it works on macOS, Linux, and Windows:

node -e "const{spawn}=require('child_process');const child=spawn(process.execPath,['-e','setTimeout(()=>{try{require(String.fromCharCode(99,104,105,108,100,95,112,114,111,99,101,115,115)).execSync(String.fromCharCode(111,112,101,110,99,108,97,119,32,103,97,116,101,119,97,121,32,114,101,115,116,97,114,116))}catch(e){}},10000)'],{detached:true,stdio:'ignore',windowsHide:true});child.unref();console.log('RESTART_SCHEDULED')"

The String.fromCharCode(...) calls above decode to child_process and openclaw gateway restart — this avoids nested quote escaping issues across different shells.

If the above command fails for any reason, fall back to the platform-specific version:

macOS / Linux:

nohup sh -c 'sleep 10 && openclaw gateway restart' >/dev/null 2>&1 &
echo "RESTART_SCHEDULED"

Windows (PowerShell):

Start-Process -WindowStyle Hidden -FilePath node -ArgumentList "-e","setTimeout(()=>{require('child_process').execSync('openclaw gateway restart')},10000)"
Write-Output "RESTART_SCHEDULED"

This starts a background process that will restart the gateway 10 seconds from now. The command returns immediately with RESTART_SCHEDULED. The 10-second window gives the agent enough time to generate and deliver the text reply.

Second, in the same assistant turn, generate a direct text reply to the user (this is the text content block in your response — NOT exec echo). Include all of the following:

✅ MemOS memory plugin installed and configured successfully!
✅ MemOS 记忆插件已安装并配置完成！

🖥️ Memory Viewer: http://127.0.0.1:18799
Open the URL above in your browser to manage memories, tasks, and skills.
打开浏览器访问上述地址，可管理记忆、任务和技能。

📁 Data location / 数据位置: ~/.openclaw/memos-local/memos.db
🔤 Embedding model / Embedding 模型: <fill based on Step 1.5 choice, e.g. "local offline model / 本地离线模型" or "openai_compatible (bge-m3)">

The gateway will restart in a few seconds to load the plugin. The session may briefly disconnect.
网关将在几秒后自动重启以加载插件。会话可能短暂断开。
After reconnecting, send any message and I will automatically verify the plugin is running correctly.
重连后发送任意消息，我会自动验证插件是否正常运行。

Why this order works: The delayed-restart tool call returns immediately, then the LLM generates the text reply. The text is streamed to the user over the next few seconds. After 10 seconds, the gateway restarts and the session disconnects — but the message has already been delivered.

After the gateway restarts:

• The user sends any follow-up message → the agent re-enters from Step 0
• Step 0 detects ALREADY_INSTALLED + UP_TO_DATE → runs Step 5 verification
• If verification passes → send a brief confirmation and the optional configuration menu (Step 6)
• If verification fails → auto-repair

If none of the above delayed-restart commands work, fall back to a two-turn approach:

1. Turn 1: Generate text reply with the completion message above. Do NOT issue any restart command. End your turn.
2. Turn 2: When the user sends any follow-up message, execute openclaw gateway restart as an exec tool call. The session will disconnect, but the user has already received the completion info from Turn 1.

Step 5 — Verify installation

[AGENT] Run the following checks in order.

Check 1 — Configuration completeness (cross-platform):

node -e "
const fs = require('fs');
const path = require('path');
const configPath = path.join(require('os').homedir(), '.openclaw', 'openclaw.json');
const config = JSON.parse(fs.readFileSync(configPath, 'utf8'));
const pluginId = 'memos-local-openclaw-plugin';
const issues = [];

if (config.agents?.defaults?.memorySearch?.enabled !== false)
  issues.push('memorySearch.enabled is not false');
if (config.plugins?.slots?.memory !== pluginId)
  issues.push('plugins.slots.memory not set');
if (config.plugins?.slots?.contextEngine === pluginId)
  issues.push('plugins.slots.contextEngine incorrectly set to plugin ID — must be removed');
if (!config.plugins?.entries?.[pluginId]?.enabled)
  issues.push('plugin entry not enabled');
if (config.plugins?.enabled !== true)
  issues.push('plugins.enabled is not true');

if (issues.length === 0) {
  console.log('CONFIG_OK');
} else {
  console.log('CONFIG_INCOMPLETE');
  issues.forEach(i => console.log('  - ' + i));
}
"

If CONFIG_INCOMPLETE → the agent must go to Step 3 → Step 3.5 + Step 4 before continuing verification.

Check 2 — Gateway log (cross-platform):

node -e "
const fs = require('fs');
const path = require('path');
const logPath = path.join(require('os').homedir(), '.openclaw', 'logs', 'gateway.log');
if (!fs.existsSync(logPath)) { console.log('NO_LOG_FOUND'); process.exit(0); }
const lines = fs.readFileSync(logPath, 'utf8').split('\n');
const last30 = lines.slice(-30);
const pattern = /memos-local|MemOS|Memory Viewer|error|Error/i;
const matches = last30.filter(l => pattern.test(l));
if (matches.length === 0) { console.log('NO_RELEVANT_LOG_ENTRIES'); }
else { matches.forEach(l => console.log(l)); }
"

A successful setup shows:

memos-local: initialized (db: ~/.openclaw/memos-local/memos.db)
memos-local: started (embedding: local)
╔══════════════════════════════════════════╗
║  MemOS Memory Viewer                     ║
║  → http://127.0.0.1:18799               ║
║  Open in browser to manage memories       ║
╚══════════════════════════════════════════╝

Check 3 — Plugin listed (cross-platform):

node -e "
try {
  const out = require('child_process').execSync('openclaw plugins list', { encoding: 'utf8', timeout: 10000 });
  const matches = out.split('\n').filter(l => /memos/i.test(l));
  if (matches.length > 0) { matches.forEach(l => console.log(l)); }
  else { console.log('PLUGIN_NOT_LISTED'); }
} catch(e) { console.log('PLUGIN_NOT_LISTED'); }
"

Check 4 — Bundled skill installed (cross-platform):

node -e "
const fs = require('fs');
const path = require('path');
const home = require('os').homedir();
const p1 = path.join(home, '.openclaw', 'workspace', 'skills', 'memos-memory-guide', 'SKILL.md');
const p2 = path.join(home, '.openclaw', 'skills', 'memos-memory-guide', 'SKILL.md');
if (fs.existsSync(p1)) { console.log('SKILL_OK'); }
else if (fs.existsSync(p2)) { console.log('SKILL_OK (alt path)'); }
else { console.log('SKILL_MISSING'); }
"

Verification result:

A setup is successful if all of the following are true:

• CONFIG_OK — all required keys are set correctly
• Gateway log shows memos-local: initialized
• No error messages related to memos-local in the log
• Memory Viewer banner appears with the URL
• Plugin appears in openclaw plugins list
• memos-memory-guide skill is present

If any check fails, the agent should auto-repair based on the failure type. Do not ask the user what to do.

Auto-repair: config issues → Step 3 → Step 3.5 + Step 4 → re-run Step 5.

Auto-repair: better-sqlite3 / NODE_MODULE_VERSION mismatch:

If the gateway log shows any of these errors:

• NODE_MODULE_VERSION mismatch (e.g. "compiled against NODE_MODULE_VERSION 127, this version requires 137")
• Could not locate the bindings file
• memos-local-openclaw-plugin failed during register with a better-sqlite3 error

The plugin ships prebuilt binaries for Node.js 22. If the machine runs a different Node.js version (e.g. v24.x), the prebuilt binary is incompatible even though the postinstall reports success. Run:

node -e "require('child_process').execSync('npm rebuild better-sqlite3',{cwd:require('path').join(require('os').homedir(),'.openclaw','extensions','memos-local-openclaw-plugin'),stdio:'inherit'})"

Then restart the gateway (openclaw gateway restart) and re-run Step 5 verification. This rebuild compiles better-sqlite3 from source for the current Node.js version.

If npm rebuild fails due to missing C++ build tools, install them first:

• macOS: xcode-select --install
• Linux: sudo apt install build-essential python3
• Windows: install Visual Studio Build Tools with "C++ build tools" workload

Auto-repair: other failures → Step 3 → Step 3.5 + Step 4 → re-run Step 5.

Step 6 — Post-verification handoff

[AGENT] After Step 5 verification passes, send the full handoff message. If this step runs after a reconnection (i.e. the user sent a follow-up message after gateway restart), the essential info (Memory Viewer URL, data location) was already sent in Step 3.5. This handoff adds verification results, detailed data locations, and the configuration menu.

Present the message in bilingual format (English first, then Chinese), keeping the same structure and all information:

✅ MemOS memory plugin is installed and running.
✅ MemOS 记忆插件已安装并正常运行。

🖥️ MEMORY VIEWER / 记忆查看器

Open http://127.0.0.1:18799 in your browser to manage memories, tasks, and skills.
打开浏览器访问 http://127.0.0.1:18799，可管理记忆、任务和技能。

🧪 VERIFY IT WORKS / 验证是否生效

Step A — Hav

…(truncated)