安装 C/C++ 编译工具
+
MemOS OpenClaw 插件MemOS OpenClaw PluginMemOS
++ 为 OpenClaw 提供完全本地化的持久记忆、智能任务总结、技能自动进化和多智能体协同。npm 一键安装，支持分级模型配置。 + Fully local persistent memory, smart task summarization, auto skill evolution, and multi-agent collaboration for OpenClaw. One-command install, tiered model support. +
+
+ 完全本地化：数据存于本机 SQLite，零云依赖。Viewer 仅 127.0.0.1，密码保护。
+ Fully local: Data in local SQLite, zero cloud dependency. Viewer 127.0.0.1 only, password-protected.
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+全量写入Full-Write
每次对话自动捕获，语义分片后持久化。Auto-captures every conversation, chunks semantically.
任务总结与技能进化Tasks & Skills
碎片对话归纳为结构化任务，再提炼为可复用技能并持续升级。Conversations organized into tasks, then distilled into skills that auto-upgrade.
混合检索Hybrid Search
FTS5 + 向量，RRF，MMR，时间衰减。FTS5 + vector, RRF, MMR, recency decay.
全量可视化Visualization
记忆/任务/技能/分析/日志/导入/设置 7 个管理页。7 pages: memories, tasks, skills, analytics, logs, import, settings.
分级模型Tiered Models
Embedding/摘要/技能可独立配置不同模型。Each pipeline configurable with different models.
多智能体协同Multi-Agent
记忆隔离 + 公共记忆 + 技能共享，多 Agent 协同进化。Memory isolation + public memory + skill sharing for collective evolution.
原生记忆导入Native Memory Import
一键迁移 OpenClaw 内置记忆，智能去重、断点续传、实时进度。One-click migration from OpenClaw built-in memories with smart dedup, resume, and real-time progress.
团队共享中心Team Sharing Hub
Hub-Client 架构，跨实例共享记忆/任务/技能。审批流程、角色管理、实时通知、端口自动推导。Hub-Client architecture for cross-instance sharing. Approval flow, role management, real-time notifications, auto port derivation.
LLM 智能降级LLM Fallback Chain
技能模型 → 摘要模型 → OpenClaw 原生模型三级自动降级，零手动干预。Skill model → summarizer → OpenClaw native model, auto-fallback with zero manual intervention.
任务/技能 CRUDTask & Skill CRUD
列表卡片直接编辑、删除、重试技能生成、切换可见性。Edit, delete, retry skill gen, toggle visibility — all from list cards.
系统架构Architecture
+四条流水线：记忆写入 → 任务总结与技能进化（异步）→ 智能检索 → 协同共享。每个 Agent 拥有独立记忆空间，通过公共记忆和技能共享实现协同进化。Four pipelines: write → task & skill evolution (async) → retrieval → collaboration. Each agent has isolated memory; public memory and skill sharing enable collective evolution.
+ +
+
OpenClawagent_end
→
+ Capture
→
+ Ingestchunk→summary→embed→dedup
→
+ SQLite+FTS5
+
+
Task Processor异步 · 话题检测 → 摘要async · topic → summary
→
+ Skill Evolver异步 · 评估 → 生成/升级async · eval → create/up
+
+
before_agent_startauto-recall
→
+ RecallFTS+Vector
→
+ LLM filter
→
+ Inject context
+
+
Agentmemory_search
→
+ RRF→MMR→Decay
→
+ LLM filter
→
+ excerpts+chunkId/task_id
→
+ task_summary / skill_get / memory_timeline
+数据流Data Flow
+写入Write
+-
+
agent_end→ Capture → Chunk → LLM Summary → Embed → Dedup → Store
+ - 异步：任务检测 → 任务摘要 → 技能评估 → 技能生成/升级Async: task detect → summary → skill eval → create/upgrade +
检索Read
+-
+
- 每轮自动：
before_agent_start用用户消息检索 → LLM 过滤相关 → 注入 system 上下文；无结果时提示 agent 自生成 query 调memory_search。Per turn:before_agent_startsearches with user message → LLM filters relevant → inject system context; if no hits, hint agent to callmemory_searchwith self-generated query.
+ memory_search→ FTS5+Vector → RRF → MMR → Decay → LLM filter → excerpts + chunkId/task_id（无 summary）
+ task_summary/skill_get(skillId|taskId) /memory_timeline(chunkId) /skill_install
+
快速开始Quick Start
+-
+
- Node.js ≥ 18 +
- OpenClaw 已安装OpenClaw installed +
- Embedding / Summarizer API 可选，不配自动用本地模型Embedding / Summarizer APIs optional, falls back to local +
Step 0：安装 C++ 编译工具（macOS / Linux 推荐）Step 0: Install C++ Build Tools (macOS / Linux recommended)
+插件依赖 better-sqlite3 原生模块。macOS / Linux 用户建议先安装编译工具，可大幅提升安装成功率。Windows 用户使用 Node.js LTS 版本时通常有预编译文件，可直接跳到 Step 1。The plugin depends on better-sqlite3, a native C/C++ module. macOS / Linux users should install build tools first. Windows users with Node.js LTS usually have prebuilt binaries and can skip to Step 1.
# macOS
+xcode-select --install
+
+# Linux (Ubuntu / Debian)
+sudo apt install build-essential python3
+
+# Windows: 通常无需操作。如安装失败，安装 Visual Studio Build Tools:
+# https://visualstudio.microsoft.com/visual-cpp-build-tools/bashStep 1：安装插件 & 启动Step 1: Install Plugin & Start
+openclaw plugins install @memtensor/memos-local-openclaw-plugin
+openclaw gateway startbash安装失败？最常见的问题是
+
+better-sqlite3 原生模块编译失败。请确认已执行上方 Step 0，然后手动重建：cd ~/.openclaw/extensions/memos-local-openclaw-plugin && npm rebuild better-sqlite3。更多方案请查看 安装排查指南 或 better-sqlite3 官方文档。Install failed? The most common issue is better-sqlite3 compilation failure. Ensure Step 0 is done, then manually rebuild: cd ~/.openclaw/extensions/memos-local-openclaw-plugin && npm rebuild better-sqlite3. See the troubleshooting guide or official better-sqlite3 docs for more solutions.升级Upgrade
+openclaw plugins update memos-local-openclaw-plugin
+openclaw gateway stop && openclaw gateway startbash升级自动完成依赖安装、旧版清理和原生模块编译，无需手动操作。如果 update 命令不可用，先删除旧目录再重新安装：
+
+rm -rf ~/.openclaw/extensions/memos-local-openclaw-plugin && openclaw plugins install @memtensor/memos-local-openclaw-plugin（记忆数据不受影响）。Upgrade automatically handles dependencies, legacy cleanup, and native module compilation. If update is unavailable, delete the old directory first: rm -rf ~/.openclaw/extensions/memos-local-openclaw-plugin && openclaw plugins install @memtensor/memos-local-openclaw-plugin (memory data is stored separately and won't be affected).配置Configuration
+两种方式：编辑 openclaw.json 或通过 Viewer 网页面板在线修改。支持分级模型。Two methods: edit openclaw.json or via Viewer web panel. Tiered models supported.
{
+ "plugins": {
+ "slots": { "memory": "memos-local-openclaw-plugin" },
+ "entries": { "memos-local-openclaw-plugin": {
+ "config": {
+ "embedding": { // lightweight
+ "provider": "openai_compatible",
+ "model": "bge-m3",
+ "endpoint": "https://your-api-endpoint/v1",
+ "apiKey": "sk-••••••"
+ },
+ "summarizer": { // mid-tier
+ "provider": "openai_compatible",
+ "model": "gpt-4o-mini",
+ "endpoint": "https://your-api-endpoint/v1",
+ "apiKey": "sk-••••••"
+ },
+ "skillEvolution": {
+ "summarizer": { // high-quality
+ "provider": "openai_compatible",
+ "model": "claude-4.6-opus",
+ "endpoint": "https://your-api-endpoint/v1",
+ "apiKey": "sk-••••••"
+ }
+ },
+ "recall": { // optional
+ "vectorSearchMaxChunks": 0 // 0=search all; set 200000–300000 only if slow on huge DB
+ },
+ "viewerPort": 18799
+ }
+ }}
+ }
+}json安装后每次对话自动存入记忆。访问
+http://127.0.0.1:18799 使用 Viewer。Every conversation auto-stored. Visit http://127.0.0.1:18799 for Viewer.🦐 记忆迁移 — 再续前缘🦐 Memory Migration — Reconnect
+将 OpenClaw 原生内置的记忆数据（SQLite 存储的对话历史）无缝迁移到 MemOS 的智能记忆系统。你和 AI 共同积累的每一段对话，都值得被记住。Seamlessly migrate OpenClaw's native built-in memory data (SQLite conversation history) to MemOS's intelligent memory system. Every conversation you've built with AI deserves to be remembered.
+ +核心特性：一键导入 · 智能去重 · 断点续传 · 任务与技能生成 · 实时进度 · 🦐 标识导入来源Key Features: One-click import · Smart dedup · Resume anytime · Task & skill gen · Real-time progress · 🦐 source tagging
+
+操作步骤Usage
+方式一：通过 Viewer 网页面板（推荐）Method 1: Via Viewer Web Panel (Recommended)
+-
+
- 访问
http://127.0.0.1:18799，切换到 Import 页面。Visithttp://127.0.0.1:18799, switch to the Import page.
+ - 点击 扫描 OpenClaw 原生记忆，系统自动扫描
~/.openclaw/下的 SQLite 数据库和 JSONL 日志。Click Scan OpenClaw Native Memories — the system auto-scans SQLite databases and JSONL logs under~/.openclaw/.
+ - 查看扫描结果（文件数、会话数、消息数），确认后点击 开始导入。Review scan results (files, sessions, messages), then click Start Import. +
- 实时查看导入进度条、统计数据（已导入/跳过/合并/错误）和日志。Monitor real-time progress bar, stats (stored/skipped/merged/errors), and logs. +
方式二：通过 Agent 对话Method 2: Via Agent Chat
+在与 OpenClaw 的对话中，直接让 AI 操作：In your conversation with OpenClaw, tell the AI:
+// Example prompts
+"请帮我导入 OpenClaw 的原生记忆"
+"Import my OpenClaw native memories"text方式三：通过 HTTP APIMethod 3: Via HTTP API
+# 1. 扫描
+curl http://127.0.0.1:18799/api/migrate/scan
+
+# 2. 开始导入（SSE 流式进度）
+curl http://127.0.0.1:18799/api/migrate/start
+
+# 3. 停止导入
+curl -X POST http://127.0.0.1:18799/api/migrate/stopbash后处理：任务与技能生成Post-Processing: Task & Skill Generation
+导入完成后，可选择对导入的记忆进行后处理：After import, optionally post-process imported memories:
+-
+
- 任务生成：自动检测会话中的任务边界，为每个会话生成结构化摘要（目标/步骤/结果）。Task generation: Auto-detect task boundaries per session, generate structured summaries (goal/steps/result). +
- 技能进化：从已完成的任务中提炼可复用技能，生成 SKILL.md 文件并安装到工作区。Skill evolution: Distill reusable skills from completed tasks, generate SKILL.md and install to workspace. +
后处理在同一 Agent 内串行执行，不同 Agent 之间可并行（并发度可配置 1–8）。已处理过的会话自动跳过。支持选择只生成任务、只生成技能或两者同时执行。Post-processing runs serially within each agent, with parallel processing across agents (configurable concurrency 1–8). Already processed sessions are auto-skipped. Choose task-only, skill-only, or both.
+ +断点续传Resume & Stop
+导入和后处理均支持随时暂停：Both import and post-processing support pause/resume:
+-
+
- 点击 停止 按钮后，进度自动保存。Click Stop, progress auto-saved. +
- 刷新页面后自动检测未完成的导入，恢复进度条显示。On page refresh, auto-detect incomplete imports and restore progress display. +
- 再次点击开始即从上次中断处继续，已处理的记忆自动跳过。Click start again to continue from where you left off — processed memories are auto-skipped. +
- 导入和后处理在后台运行，关闭 Viewer 页面不影响执行。Import and post-processing run in the background — closing the Viewer page won't interrupt them. +
🦐 来源标识：所有通过迁移导入的记忆都带有 🦐 标识，在 Viewer 的记忆列表中可一眼区分原生导入和对话生成的记忆。🦐 Source Tag: All migrated memories are tagged with 🦐, making them visually distinguishable from conversation-generated memories in the Viewer.
+模块Modules
+Capture
+过滤 system/self-tool，剥离 OpenClaw 元数据。保留 user/assistant/tool。Filter system/self-tool, strip metadata. Keep user/assistant/tool.
+Ingest
+异步队列：语义分片 → LLM 摘要 → 向量化 → 智能去重（Top-5 相似 + LLM 判 DUPLICATE/UPDATE/NEW，UPDATE 合并摘要并追加内容）→ 存储；演化块记录 merge_history。Async queue: chunk → summary → embed → smart dedup (Top-5 similar + LLM DUPLICATE/UPDATE/NEW; UPDATE merges summary and appends content) → store; evolved chunks track merge_history.
+任务总结Task Summarization
+异步逐轮检测任务边界：分组为用户回合 → 第一条直接分配 → 后续每条由 LLM 判断话题是否切换（强偏向 SAME，避免过度分割）→ 2h 超时强制切分 → 结构化摘要（目标/步骤/结果）。支持编辑、删除、重试技能生成。Async per-turn boundary detection: group into user turns → first turn assigned directly → each subsequent turn checked by LLM topic judge (strongly biased toward SAME to avoid over-splitting) → 2h timeout forces split → structured summary (goal/steps/result). Supports edit, delete, retry skill generation.
+技能进化Skill Evolution
+规则过滤 → LLM 评估（可重复/有价值的任务才生成技能）→ SKILL.md 生成（步骤/警告/脚本）/ 升级 → 质量评分 → 安装。LLM 使用三级降级链（技能模型 → 摘要模型 → OpenClaw 原生模型）。支持编辑、删除、设为公开/私有。Rule filter → LLM evaluate (only repeatable/valuable tasks generate skills) → SKILL.md (steps/warnings/scripts) / upgrade → score → install. LLM uses a 3-level fallback chain (skill model → summarizer → OpenClaw native model). Supports edit, delete, toggle visibility.
+Recall
+FTS5+Vector → RRF(k=60) → MMR(λ=0.7) → Decay(14d) → Normalize → Filter(≥0.45) → Top-K。自动关联 Task/Skill。FTS5+Vector → RRF(k=60) → MMR(λ=0.7) → Decay(14d) → Normalize → Filter(≥0.45) → Top-K. Auto-links Task/Skill.
+Viewer
+7 页：记忆 CRUD/搜索/演化标识、任务（对话气泡）、技能（版本/下载）、分析、日志（工具调用输入输出）、OpenClaw 原生记忆导入、在线配置。密码保护。7 pages: memory CRUD/search/evolution badges, tasks (chat bubbles), skills (versions/download), analytics, logs (tool call I/O), OpenClaw native memory import, online config. Password-protected.
+检索算法Retrieval
+RRF
+\[ \text{RRF}(d) = \sum_i \frac{1}{k + \text{rank}_i(d) + 1} \]
+MMR
+\[ \text{MMR}(d) = \lambda \cdot \text{rel}(d) - (1-\lambda) \cdot \max \text{sim}(d, d_s) \]
+时间衰减Recency
+\[ \text{final} = \text{score} \times \bigl(0.3 + 0.7 \times 0.5^{t/14}\bigr) \]
+API
+memory_search
+query (required), maxResults (20), minScore (0.45), role. Returns excerpts（原文片段）+ chunkId / task_id，无 summary；经 LLM 相关性过滤。excerpts + chunkId/task_id, no summary; LLM relevance filter.
memory_get
+获取记忆块完整原文。Get full original text of a memory chunk. chunkId, maxChars (optional).
memory_timeline
+以 chunkId 为锚点的上下文邻居。Context neighbors by chunkId. chunkId, window (2).
task_summary
+任务结构化摘要。Structured task summary. taskId or query.
+skill_get / skill_install
+skill_get 支持 skillId 或 taskId（按任务解析技能）；skill_install 安装到工作区。skill_get accepts skillId or taskId; skill_install installs to workspace.
+memory_write_public
+写入公共记忆（owner="public"），所有 Agent 均可检索。Write public memory (owner="public"), discoverable by all agents. content (required), summary (optional).
skill_search
+搜索技能：FTS5 关键词 + 向量语义双通道，RRF 融合后经 LLM 判断相关性。Search skills via FTS5 + vector, RRF fusion, then LLM relevance judgment. query (required), scope ("mix" | "self" | "public", default "mix").
skill_publish / skill_unpublish
+skill_publish 将技能设为公开，其他 Agent 可通过 skill_search 发现并安装。skill_unpublish 设为私有。skill_publish makes a skill public and discoverable via skill_search. skill_unpublish sets it private. skillId (required).
memory_viewer
+返回 Viewer URL。Returns Viewer URL.
+Viewer HTTP
+| Method | Path | 说明Description |
|---|---|---|
| GET | / | Memory Viewer HTML |
| POST | /api/auth/* | setup / login / reset / logout |
| GET | /api/memories | 记忆列表（分页、过滤）Memory list (pagination, filters) |
| GET | /api/search | 混合搜索（向量 minScore 0.64 + FTS5 降级）Hybrid search (vector minScore 0.64 + FTS5 fallback) |
| POST/PUT/DELETE | /api/memory/:id | 记忆 CRUDMemory CRUD |
| GET | /api/tasks | 任务列表（状态过滤）Task list (status filter) |
| GET/PUT/DELETE | /api/task/:id | 任务详情/编辑/删除Task detail/edit/delete |
| POST | /api/task/:id/retry-skill | 重试技能生成Retry skill generation |
| GET | /api/skills | 技能列表Skill list |
| GET/PUT/DELETE | /api/skill/:id | 技能详情/编辑/删除Skill detail/edit/delete |
| PUT | /api/skill/:id/visibility | 设置公开/私有Set public/private |
| GET | /api/skill/:id/download | 技能 ZIP 下载Download as ZIP |
| GET | /api/stats, /api/metrics | 统计与分析Stats & metrics |
| GET | /api/logs | 工具调用日志Tool call logs |
| GET/PUT | /api/config | 在线配置Online configuration |
| GET/POST | /api/migrate/* | 记忆导入（扫描/开始/停止/SSE 进度）Memory import (scan/start/stop/SSE) |
| POST/GET | /api/migrate/postprocess/* | 后处理（任务/技能生成）Post-process (task/skill gen) |
多智能体协同Multi-Agent Collaboration
+MemOS 原生支持多 Agent 场景。每个 Agent 的记忆和任务通过 owner 字段隔离（格式 agent:{agentId}），检索时自动过滤为当前 Agent + public。MemOS natively supports multi-agent scenarios. Each agent's memories and tasks are isolated via an owner field (agent:{agentId}); retrieval automatically filters to current agent + public.
-
+
- 记忆隔离：Agent A 无法检索 Agent B 的私有记忆Memory Isolation: Agent A cannot retrieve Agent B's private memories +
- 公共记忆：通过
memory_write_public写入 owner="public" 的记忆，所有 Agent 可检索Public Memory: Usememory_write_publicto write owner="public" memories discoverable by all agents
+ - 技能共享：通过
skill_publish将技能设为公开，其他 Agent 可通过skill_search发现并安装Skill Sharing: Useskill_publishto make skills public; other agents discover and install viaskill_search
+ - 技能检索：
skill_search支持 scope 参数（mix/self/public），FTS + 向量双通道 + RRF 融合 + LLM 相关性判断Skill Discovery:skill_searchsupports scope (mix/self/public), FTS + vector dual channel + RRF fusion + LLM relevance judgment
+
LLM 降级链LLM Fallback Chain
+所有 LLM 调用（摘要、话题检测、去重、技能生成/升级）均使用三级自动降级机制：All LLM calls (summary, topic detection, dedup, skill generation/upgrade) use a 3-level automatic fallback chain:
+
+
skillSummarizer技能专用模型（可选）Skill-dedicated (optional)
→
+ summarizer通用摘要模型General summarizer
→
+ OpenClaw Native从 openclaw.json 读取Auto-detected from openclaw.json
+-
+
- 每一级失败后自动尝试下一级，无需手动干预Each level auto-falls back to the next on failure, zero manual intervention +
skillSummarizer未配置时直接跳到summarizerIfskillSummarizeris not configured, skips directly tosummarizer
+ - OpenClaw 原生模型从
~/.openclaw/openclaw.json的agents.defaults.model.primary自动读取OpenClaw native model auto-detected from~/.openclaw/openclaw.json→agents.defaults.model.primary
+ - 如果所有模型均失败，回退到规则方法（无 LLM）或跳过该步骤If all models fail, falls back to rule-based methods (no LLM) or skips the step +
数据库Database
+~/.openclaw/memos-local/memos.db, WAL. Tables: chunks (owner), chunks_fts, embeddings, tasks (owner), skills (owner, visibility), skill_versions, task_skills, skill_embeddings, skills_fts.
安全Security
+Viewer 仅 127.0.0.1；密码 SHA-256；HttpOnly+SameSite Cookie；会话 24h；数据仅本地。127.0.0.1 only; SHA-256 password; HttpOnly+SameSite; 24h session; data stays local.
+默认值Defaults
+| 参数Parameter | 默认Default | 说明Description |
|---|---|---|
| maxResults | 6 (max 20) | 默认返回数Default result count |
| minScore (tool) | 0.45 | memory_search 最低分memory_search minimum |
| minScore (viewer) | 0.64 | Viewer 搜索向量阈值Viewer search vector threshold |
| rrfK | 60 | RRF 融合常数RRF fusion constant |
| mmrLambda | 0.7 | MMR 相关性 vs 多样性MMR relevance vs diversity |
| recencyHalfLife | 14d | 时间衰减半衰期Recency decay half-life |
| vectorSearchMaxChunks | 0 (all) | 0=搜索全部；大库可设 200k-300k0=search all; set 200k-300k for large DBs |
| dedup threshold | 0.75 | 语义去重余弦相似度Semantic dedup cosine similarity |
| viewerPort | 18799 | Memory Viewer |
| taskIdle | 2h | 任务空闲超时Task idle timeout |
| topicJudgeWarmup | 1 | LLM 话题判断预热（用户消息数）LLM topic judge warm-up (user turns) |
| skillMinChunks | 6 | 技能评估最小 chunk 数Min chunks for skill evaluation |
| importConcurrency | 1 (max 8) | 导入 Agent 并行度Import agent parallelism |