[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-80202":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":16,"forks30d":16,"starsTrendScore":19,"compositeScore":20,"rankGlobal":10,"rankLanguage":10,"license":21,"archived":22,"fork":22,"defaultBranch":23,"hasWiki":22,"hasPages":22,"topics":24,"createdAt":10,"pushedAt":10,"updatedAt":40,"readmeContent":41,"aiSummary":42,"trendingCount":16,"starSnapshotCount":16,"syncStatus":15,"lastSyncTime":43,"discoverSource":44},80202,"overmind","xuanlinAI\u002Fovermind","xuanlinAI","玄霖超脑 · 无量网络 v4 重构版— 66 模块6通道 AI 认知神经系统 · 装一次，你所有的 AI 工具从此共享一个永远不失忆的大脑。跨会话记忆 · 多 Agent 互通 · 自动代码审查 · 零配置","",null,"JavaScript",82,10,73,2,0,5,9,1,3.12,"MIT License",false,"main",[25,26,27,28,29,30,31,32,33,34,35,36,37,38,39],"ai-agent","ai-memory","codex","deepseek","hermes","knowledge-graph","llm","mcp-server","memory-management","openclaw","python","reasoning-engine","self-evolving","skill-learning","typescript","2026-06-12 02:03:59","# 玄霖超脑 · 无量网络 v4 重构版\n\n\u003Cp align=\"center\">\n  \u003Cb>首个开源的事件驱动 AI Agent 认知架构。目前最强的是 v4。未来最强的是 v5。\u003C\u002Fb>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  不是记忆，是推理。不是工具，是神经系统。不是插件，是认知引擎。\u003Cbr>\n  全球唯一 6 通道架构 · 66 模块 · 37 阶段管道 · 知识图谱 5000+ 节点 · 跨 Agent 舰队协同\u003Cbr>\n  装完，你所有的 AI 工具从此共享一个永远不失忆的大脑。\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"#中文\">🇨🇳 中文\u003C\u002Fa> | \u003Ca href=\"#english\">English\u003C\u002Fa>\n\u003C\u002Fp>\n\n---\n\n\u003Cspan id=\"中文\">\u003C\u002Fspan>\n\n# 🇨🇳 中文\n\n## 你每天都在重复这三件事\n\n1. **开一个新的 Claude Code 会话，又花二十分钟把项目从头介绍一遍。** 上次教过它的踩坑、规范、决策，全部清零。\n2. **开第二个 CC 窗口处理另一个任务，两个窗口互相不知道对方在干嘛。** 经常打架、改坏对方的代码。\n3. **CC 一脸自信地给你一段代码，跑起来才发现是幻觉。** 调试半小时才发现是它编的 API。\n\n玄霖超脑就是来解决这三件事的。\n\n## 它到底是个什么东西\n\n一句话：**Claude Code 的外挂大脑**。但不是普通大脑——是全球首个**事件驱动 AI Agent 认知架构**。它不只是存记忆，它在推理。\n\n装上之后：\n\n- 🧠 **永久记忆**：你跟 CC 说过的每一句话、做过的每一个决定、踩过的每一个坑，它都记得。下次开会话直接接着干，不用再介绍项目。\n- 🔗 **多窗口舰队共享**：开五个 CC 同时干活，它们互相能看到对方在改什么、学到什么。一个发现的 bug，所有窗口都知道。一个 CC 学到的技能，另一个 CC 直接能用。\n- 🛡️ **回答前自检**：CC 给出代码之前，超脑先在后台过一遍——这段代码会不会和已有架构冲突？这个 API 是不是它编的？有问题先拦下来重写，再给你看。\n- 🎯 **越用越懂你**：你的代码风格、命名习惯、架构偏好，它会自己总结。一周后它写的代码，会越来越像你写的。\n\n## 凭什么说是最强的\n\n**没有竞品能做到这些：**\n\n| | 超脑 v4 | Mem0 | AgentMemory | Cursor | Copilot |\n|---|---|---|---|---|---|\n| 事件驱动架构 | ✅ 6通道 | ❌ | ❌ | ❌ | ❌ |\n| 跨会话永久记忆 | ✅ | ⚠️ | ⚠️ | ❌ | ❌ |\n| 多 Agent 舰队协同 | ✅ 实时 | ❌ | ❌ | ❌ | ❌ |\n| 回答前自检拦截 | ✅ 默认 | ❌ | ❌ | ❌ | ❌ |\n| 代码风格自动学习 | ✅ | ❌ | ❌ | ⚠️ | ❌ |\n| 知识图谱推理 (10种关系) | ✅ 5000+节点 | ❌ | ❌ | ❌ | ❌ |\n| 零配置自适应安装 | ✅ \u003C1分钟 | ❌ | ❌ | - | - |\n| 全本地 | ✅ | ❌ 上云 | ✅ | ❌ 上云 | ❌ 上云 |\n| 免费 | ✅ MIT | ❌ $249\u002F月 | ✅ | ❌ $20\u002F月 | ❌ $10\u002F月 |\n\n**一句话：他们做的是 AI 的硬盘。Overmind 做的是 AI 的脑。不是一个维度的东西。**\n\n## 30 秒安装\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FxuanlinAI\u002Fovermind.git && cd overmind && node install.js\n```\n\n安装脚本会交互式问你 API 配置（API 端点、Key、模型 ID），按回车用默认 DeepSeek。装完打开 Claude Code，超脑已经在背后跑了。\n\n```bash\nexport OVERMIND_API_KEY=sk-xxx        # 推荐：超脑专用（也支持 DEEPSEEK_API_KEY \u002F ANTHROPIC_AUTH_TOKEN）\n```\n\n## 66 个模块到底在干什么\n\n### 记忆系统 (5 层)\n语义记忆 (SQLite + FTS5 + jieba 中文分词，AI 精选) · 程序性记忆 (同模式 ≥3 次自动晋升) · 情景记忆 (AI 生成摘要，按项目检索) · 知识图谱 (10 种关系，BFS 被动推理) · 反馈回路 (4500+ 事件驱动有效性自动调节)\n\n### 37 阶段管道 (CH1)\n意图预判 · 安全门禁 (rm -rf 拦截) · 提交门禁 (未 commit 先挡) · 输出盾 (过期 API 检测) · **人格画像 (8维度推断，自动boost技能)** · **红队审查 (恶魔分身攻击每个输出)** · 异常检测 (领域偏移\u002F记忆暴涨) · 预测调试 (git diff → 历史失败) · 成本分析 · 跨会话续接 · **技能编排 (检测 A→B→C 链，自动生成 meta-skill)** · **反事实引擎 (决策阴影追踪)** · 知识验证 · 预测调试 · 检查点回滚 · TDD 约束 · 环境预取 · 噪声学习 · **心理模型 (盲点\u002F重复错误\u002F技能缺口)** · **梦境研究 (Pro 模型深度分析，你不在它也在思考)** · 创意合成 · 思维链透明 · 自主研究 · 预算拦截 (3次失败→止损) · 跨项目迁移 (6领域) · 技能血统 · 记忆预算 · 反流失 · 预判加载 · 会话简报 · 因果可视化 · **时间旅行 (回到任意日期\u002Fcommit 的记忆状态)** · 僵尸检测 (90天未访问) · 自愈检查 · 模块链接\n\n### z2 中枢 (CH3+CH4)\n**FleetWatcher (5秒扫描所有 CC 会话，提取 Q\u002FA，生成舰队广播)** · CH3 直连 (!include → LLM 自动感知同伴) · CH4 中枢广播 (event_queue → bus → 9模块) · 零进程触发 (根治 Windows 窗口闪烁) · 20 个 MCP 工具\n\n### n2 终端 (CH5+CH6)\nAI 筛选 (94% 压缩率) · CH5 8 模块串联补全 · CH6 11 模块并联通知\n\n### 舰队 + 涌现功能\n**多 CC 自发现 (自动感知所有 CC 实例)** · 跨 CC 技能 boosting (同伴做逆向→自动 boost 逆向技能) · 舰队冲突检测 (两个 CC 同文件→告警) · 跨 CC 创意碰撞 (不同域工作→交叉洞察) · **自我镜像回路 (广播含自身对话)** · 技能自动贴标签 (30\u002F30)\n\n### 自进化\nHermes Fusion (14轮) · 自动晋升 (70次) · AI 工作流 (4个) · 技能草案 (30个 Worker 自动创建) · 记忆压缩 (2145条)\n\n### 安全\n红队审查 · 安全拦截 · 输出盾 · 提交门禁 · TDD 约束 · 预算拦截 · 补丁模式\n\n## 装完之后发生什么\n\n```\n你：继续昨天那个登录模块\nCC：好的，昨天我们用了 JWT + Redis 黑名单，refresh token 设了 7 天。\n    现在卡在 SSO 跳转那块。继续从那里开始？\n```\n不需要重新介绍项目。\n\n```\n窗口 A：在改 user.service.ts\n窗口 B：（自动提示）窗口 A 正在改 user.service.ts 的登录逻辑，我先避开这个文件。\n```\n开五个 CC 同时干活，不会打架。\n\n```\n[超脑后台：检查中... fetchUserProfile 这个函数项目里不存在，\n            CC 可能是幻觉。打回重写。]\nCC：（重新生成）这是基于你项目里实际存在的 getUserById 写的版本……\n```\n幻觉还没到你眼前就被拦下了。\n\n## 重构版 vs 初版\n\n| 维度 | v4 初版 | v4 重构版 |\n|------|---------|-----------|\n| 启动速度 | 8-12 秒 | \u003C 2 秒 |\n| 内存占用 | 600-900 MB | 120-200 MB |\n| 模块数 | 66 个，互相耦合 | 精简到核心层 + 插件层 |\n| 配置复杂度 | 改 5-7 个文件 | 一条命令 |\n| 多实例同步 | 文件锁，会卡 | SQLite WAL，无感 |\n| 自检机制 | 串行，慢 | 并行 + 提前终止，快 5-8 倍 |\n| 错误恢复 | 出错就挂 | 自动降级，单模块崩溃不影响整体 |\n| 安装器 | 无 | probe+vault+config+ready 四件套，失败自动快照回滚 |\n\n## 全功能基准测试\n\n| 指标 | 数值 |\n|------|------|\n| 管道热吞吐 | 50,000 ops\u002Fs |\n| 记忆写入 | 200,000 ops\u002Fs |\n| FTS5 搜索 | 0.22ms\u002Fop |\n| 图谱边插入 | 142,857 ops\u002Fs |\n| 事件 I\u002FO | 2,342 write\u002Fs · 2,649 drain\u002Fs |\n| CH5 终端串联 | 102ms\u002Fop |\n| CH6 终端并联 | 0.78ms\u002Fop |\n| 地狱压力测试 | 5\u002F5 suites · 96\u002F100 HELL_PASSED |\n| 18 项安全审计 | 18\u002F18 PASSED |\n| 32 项 Opus bug 排查 | 8 个确认并修复 |\n| 42 项发布自测 | 42\u002F42 PASSED |\n\n## 版本演进\n\n| 版本 | 时间 | 关键变化 |\n|------|------|---------|\n| v1 | 2024 Q3 | 概念验证，66 模块全量铺开 |\n| v2 | 2024 Q4 | 加入跨会话记忆，开始有用 |\n| v3 | (跳过发布) | 30 模块，5 层记忆，18 MCP 工具 |\n| **v4 初版** | 2025 Q2 | 事件总线 · 红队 · 反事实 · 心理模型 |\n| **v4 重构版** | 2025 Q2 | 6 通道 · z2 中枢 · n2 终端 · 自适应安装 · 零进程触发 |\n| v5 | 规划中 | 中性 IR · 共享记忆图谱 · 环形编排器 · 插件总线 |\n\n## FAQ\n\n**Q：会上传我的代码吗？** 不会。所有记忆数据存在本地 SQLite。只有你调用 LLM 时才会发请求。\n\n**Q：API 费用会暴涨吗？** 自检会多消耗 token，实测增加 15-25%。但因为减少了幻觉返工，总账反而省。\n\n**Q：怎么卸载？** `node install.js --uninstall` 干净卸载。\n\n---\n\n\u003Cspan id=\"english\">\u003C\u002Fspan>\n\n# 🇬🇧 English\n\n## Three things you do every single day\n\n1. **Open a fresh Claude Code session and spend 20 minutes re-introducing your project.** Every quirk, convention, and hard-won decision from last time? Gone.\n2. **Open a second CC window for a parallel task — neither window knows what the other is doing.** They step on each other's commits.\n3. **CC confidently hands you code that turns out to be hallucinated.** Thirty minutes of debugging later, you discover it invented an API.\n\nXuanlin Overmind exists to fix exactly these three things.\n\n## What it actually is\n\nOne line: **an external brain for Claude Code.** But not just any brain — the world's first **event-driven AI agent cognitive architecture.** It doesn't just store memories. It reasons about them.\n\n- 🧠 **Permanent memory.** Every conversation, decision, and gotcha you've taught CC stays. Next session picks up where the last one ended.\n- 🔗 **Fleet sharing.** Run five CC instances in parallel. They all see what the others are touching, learning, and breaking. A bug discovered in one is known to all. A skill learned by one is usable by all.\n- 🛡️ **Pre-answer self-check.** Before CC shows you code, Overmind checks it — does this conflict with existing architecture? Is this API real or hallucinated? Catches issues before they reach you.\n- 🎯 **Learns your style.** Your naming patterns, architectural taste, error-handling habits. After a week, CC's output starts looking like code you wrote.\n\n## Why it's the strongest\n\n**No competitor can do this:**\n\n| | Overmind v4 | Mem0 | AgentMemory | Cursor | Copilot |\n|---|---|---|---|---|---|\n| Event-driven architecture | ✅ 6-channel | ❌ | ❌ | ❌ | ❌ |\n| Cross-session permanent memory | ✅ | ⚠️ | ⚠️ | ❌ | ❌ |\n| Multi-agent fleet collaboration | ✅ Real-time | ❌ | ❌ | ❌ | ❌ |\n| Pre-answer self-check | ✅ Default | ❌ | ❌ | ❌ | ❌ |\n| Code-style auto-learning | ✅ | ❌ | ❌ | ⚠️ | ❌ |\n| Knowledge graph (10 relations) | ✅ 5000+ nodes | ❌ | ❌ | ❌ | ❌ |\n| Zero-config adaptive install | ✅ \u003C1 min | ❌ | ❌ | - | - |\n| Fully local | ✅ | ❌ Cloud | ✅ | ❌ Cloud | ❌ Cloud |\n| Free | ✅ MIT | ❌ $249\u002Fmo | ✅ | ❌ $20\u002Fmo | ❌ $10\u002Fmo |\n\n**They built AI's hard drive. Overmind is AI's brain. Not the same league.**\n\n## 30-second install\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FxuanlinAI\u002Fovermind.git && cd overmind && node install.js\n```\n\nThe installer asks a few questions (API provider, key, model) and finishes in under a minute. Open any Claude Code session and Overmind is already running.\n\n```bash\nexport OVERMIND_API_KEY=sk-xxx        # Recommended (also supports DEEPSEEK_API_KEY \u002F ANTHROPIC_AUTH_TOKEN)\n```\n\n## What the 66 modules actually do\n\n### Memory System (5 layers)\nSemantic (SQLite+FTS5+jieba, AI-curated) · Procedural (≥3 patterns auto-promoted) · Episodic (AI-generated summaries) · Knowledge Graph (10 relations, BFS reasoning) · Feedback Loop (4500+ events, auto-scoring)\n\n### 37-Stage Pipeline (CH1)\nIntent prediction · Security gate (rm -rf interception) · Commit gate · Output shield · **Persona profiling (8-dim, auto skill boost)** · **Red-team adversarial (shadow attacks every output)** · Anomaly detection · Forecast debugging · Cost optimization · Cross-session continuity · **Skill composition (A→B→C chains, auto meta-skill)** · **Counterfactual engine (decision shadow tracking)** · Verification · Predictive debugging · Checkpoints · TDD enforcement · Prefetch · Noise learning · **Theory of mind (blind spots\u002Frecurring errors)** · **Dream research (pro model analysis while idle)** · Creative synthesis · Reasoning transparency · Autonomous research · Budget interception (3 failures=stop) · Cross-project transfer (6 domains) · Skill lineage · Memory budget · Anti-compaction · Preloading · Briefing · Causal visualization · **Time travel (any date\u002Fcommit memory state)** · Zombie detection (90-day) · Self-healing · Nexus linking\n\n### z2 Hub (CH3+CH4)\n**FleetWatcher (5s scan of all CC sessions, Q\u002FA extraction, fleet broadcast)** · CH3 Direct (!include → LLM peer awareness) · CH4 Hub broadcast (event_queue → bus → 9 modules) · Zero-process trigger (eliminates Windows flash) · 20 MCP tools\n\n### n2 Terminal (CH5+CH6)\nAI filter (94% compression) · CH5 8-module serial · CH6 11-module parallel\n\n### Fleet + Emergence\n**Multi-CC auto-discovery** · Cross-CC skill boosting · Fleet conflict detection · Cross-CC creative collision · **Self-mirror loop** · Auto skill tagging (30\u002F30)\n\n### Self-evolution\nHermes Fusion (14 rounds) · Auto-promotion (70) · AI workflows (4) · Auto-created skills (30) · Memory compaction (2145)\n\n### Security\nRed-team review · Security interception · Output shield · Commit gate · TDD enforcement · Budget interceptor · Patch mode\n\n## Performance\n\n| Metric | Value |\n|--------|-------|\n| Pipeline throughput (warm) | 50,000 ops\u002Fs |\n| Memory writes | 200,000 ops\u002Fs |\n| FTS5 search | 0.22ms\u002Fop |\n| Graph edge insert | 142,857 ops\u002Fs |\n| Event I\u002FO | 2,342 write\u002Fs · 2,649 drain\u002Fs |\n| Terminal serial | 102ms\u002Fop |\n| Terminal parallel | 0.78ms\u002Fop |\n| Hell benchmark | 5\u002F5 suites · 96\u002F100 |\n| Security audit | 18\u002F18 PASSED |\n| Opus bug audit | 8 confirmed & fixed |\n| Release self-test | 42\u002F42 PASSED |\n\n## Refactored v4 vs Original\n\n| Aspect | v4 Original | v4 Refactored |\n|--------|-------------|---------------|\n| Startup | 8-12s | \u003C2s |\n| Memory | 600-900MB | 120-200MB |\n| Config | 5-7 files | One command |\n| Multi-instance | File locks, stalls | SQLite WAL, seamless |\n| Self-check | Serial | Parallel, 5-8x faster |\n| Error recovery | Hard crash | Graceful degradation |\n| Installer | None | probe+vault+config+ready, auto-rollback |\n\n## Version History\n\n| Version | Date | Highlights |\n|---------|------|------------|\n| v1 | 2024 Q3 | Proof of concept |\n| v2 | 2024 Q4 | Cross-session memory |\n| v3 | (Skipped) | 30 modules, 5-layer memory |\n| **v4 Original** | 2025 Q2 | Event bus, Red team, Counterfactual, Theory of Mind |\n| **v4 Refactored** | 2025 Q2 | 6-channel, z2 hub, n2 terminal, adaptive install |\n| v5 | Planned | Neutral IR, shared memory graph, ring orchestrator, plugin bus |\n\n## FAQ\n\n**Q: Does it upload my code?** No. All local SQLite.\n\n**Q: Will my API bill explode?** Self-check adds 15-25% tokens. Net spend decreases due to fewer hallucination re-debugs.\n\n**Q: How to uninstall?** `node install.js --uninstall`\n\n---\n\n\u003Cp align=\"center\">\n  \u003Cb>目前最强的是 v4。未来最强的是 v5。\u003C\u002Fb>\u003Cbr>\u003Cbr>\n  Made by \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FxuanlinAI\">玄霖AI (xuanlinAI)\u003C\u002Fa>\u003Cbr>\n  Issues and PRs welcome · MIT License\n\u003C\u002Fp>\n","玄霖超脑 · 无量网络 v4 重构版是一个事件驱动的AI认知架构，旨在为用户提供一个永久记忆、多Agent协同和自动代码审查的大脑。其核心功能包括跨会话记忆共享、多窗口舰队实时协同、回答前自检拦截以及自动学习用户编码风格等。该系统基于6通道架构，集成了66个模块，支持37阶段管道处理，并构建了超过5000节点的知识图谱以增强推理能力。适用于需要长期记忆保持、多任务并行处理及高质量代码生成的开发场景中，特别适合那些希望提高工作效率同时减少重复劳动的技术团队或个人开发者使用。","2026-06-11 03:59:37","CREATED_QUERY"]