0.9.0最新未发布
更新日志
Nowledge Mem 的新功能、改进和修复。
当前版本0.9.0
0.8.72026年5月20日
Search Refresh Hotfix
0.8.7 fixes a Search Refresh issue that could leave some upgraded users waiting at the final repair step when they had large Library documents. Conversation-search migration now stays lightweight and finishes without reparsing Library files.
0.8.62026年5月15日
Knowledge Tree, Nowledge Mem FS, and Safer Search Refresh
0.8.6 introduces the preview of Knowledge Tree and Nowledge FS, and makes the 0.8.4 conversation-storage upgrade safer. Search Refresh now avoids unnecessary embedding rebuilds when only conversation search needs migration, recovers cleanly from interrupted 0.8.4/0.8.5 refreshes, and stays quiet on fresh installs with no content yet.
亮点
新增
- 知识树作为 Mem 虚拟文件系统的第一层 UI 入口上线。它会用原生预览打开 Mem 对象，而不是暴露底层实现文件。
- `nmem fs` 新增了面向 Nowledge FS 的类 shell 命令：`ls`、`cat`、`stat`、`find`、`grep`、`recall`、`write` 和 `rm`。`grep` 可以搜索记忆、对话消息和已解析的资料库文档；`cat --line --lines` 只读取需要的片段。
- `/fs/*` API 和 `mem_fs` MCP 工具让集成与 Agent 使用和应用内相同的按路径浏览入口，并明确区分后续阶段的挂载、SQL/Cypher 和来源片段能力。
改进
修复
- 如果之前在 0.8.4 或 0.8.5 中断过刷新搜索，0.8.6 可以从已保存的数据修复缺失的搜索记录，并完成升级，不再强制走完整语义重建。
- 全新安装在下载嵌入模型后，如果还没有任何内容需要刷新，就不会再提示刷新搜索。
- Tree 现在可在 Web 客户端正常使用；大型分支会分页加载，不会悄悄隐藏后续条目；原生预览也不会再因为当前空间不匹配而报错。
- 新版 OpenClaw 运行时发送较短的逐轮 `agent_end` 数据时，连续会话同步不会再丢掉有效的后续轮次。
- 在 Windows 上点击 Install & Restart 后，Mem 更不容易出现安装完成但应用没有重新打开的情况。
- 启动流程现在能识别 VPN、代理或 TUN 模式工具拦截 localhost 流量的情况，并提示用户需要绕过哪些地址。
- 使用 Responses API 的 OpenAI 兼容服务商现在能更好地适配会拒绝过小输出 token 限制的第三方端点。
0.8.52026年5月15日
Search Refresh Hotfix
0.8.5 focuses on upgrade safety after 0.8.4. Search Refresh is less likely to run out of graph memory, completed refreshes no longer look failed when old conversation-storage cleanup has to wait, and stale refresh prompts from 0.8.4 are cleared when Mem can verify the rebuild already finished.
0.8.42026年5月14日
Docker Preview and Leaner Conversation Storage
0.8.4 is mostly a refactor and reliability release. Mem moves thread messages out of Knowledge Graph storage and onto a lighter Content Store path, introduces the official Docker deployment preview with web-triggered operator-sidecar updates, and fixes a long list of Linux headless and remote-access edge cases.
亮点
- 官方 Docker 部署现在以预览形式开放，适合在无头服务器、NAS、家庭服务器和 VPS 上运行 Mem。同一份已签名的多架构镜像同时覆盖 `linux/amd64` 与 `linux/arm64`（Apple Silicon 的 Docker Desktop、Ampere/Graviton、树莓派 5 等 arm64 板卡），数据则保存在普通的 bind mount 文件夹里，可以用常规工具备份。
- 大型对话库有了更轻的底层结构。对话消息会移出知识图谱存储，但对话搜索、证据高亮、备份和恢复仍然可用。
- 自托管 Docker 服务器在执行 `nmemctl auto-update enable` 后，可以直接从网页端升级。operator sidecar 会拉取新镜像、保存升级前快照，并重建容器。
- 如果升级后 Mem 提示你刷新搜索，执行一次即可。这会准备新的对话搜索路径，并保持旧对话可用。
- 无头 Linux 和远程客户端更稳了：许可证操作可以走远程访问，容器内存限制会被正确识别，智能体卡住时会明确报错，LAN/网页端 AI Now 也能可靠启动。
新增
- 官方 Docker 部署预览：可以把 Mem 跑在你自己的服务器上，使用持久化 bind mount 数据目录，支持 `nmemctl` 管理流程、可选的域名/TLS 配置，以及专门的 `/livez` 存活探针。
- 自托管 Docker 服务器现在可以通过 operator sidecar 直接在网页里升级。用 `nmemctl auto-update enable` 启用一次，之后只要发布了新版本，在设置里点 Install 就能升级：Mem 会先把数据保存为快照，再拉取新镜像并重建容器，整个过程大约三十秒。最近 3 份升级前的快照会保留在本机，万一出问题，可以从主机上回滚到之前的状态。
- `nmemctl backup-app` 和 `restore-app` 提供应用层迁移路径，可用于从桌面版或 `.deb` 安装迁移到 Docker，也可以在不同版本的 Docker 主机之间迁移。
改进
0.8.32026年5月2日
Plugin Refactor and DeepSeek V4
0.8.3 tightens the coding-agent plugin layer and adds DeepSeek V4 support. Plugin capture now follows one clearer contract across major hosts, Codex can save real sessions, Claude Code and Cursor keep repeated captures on the same thread, and DeepSeek V4 works as a first-class model choice with safer tool-loop handling.
亮点
- 主要编程智能体插件现在共享更稳定的捕获契约。插件更新、Stop hook、文件监听和解析导入更不容易把同一段对话拆成多条 Mem 线程。
- DeepSeek V4 现在可以直接用于 AI Now、Feed、Knowledge、Graph Intelligence，以及浏览器里的 Distill / Auto-Capture。Mem 会优先推荐 `deepseek-v4-flash`，并正确识别 V4 的 100 万上下文窗口。
- 使用工具的 DeepSeek V4 智能体现在会避开严格 thinking 历史校验导致的中断。官方 DeepSeek 和 OpenAI 兼容的自定义 DeepSeek V4 端点会启用兼容保护；通过 OpenRouter 使用的模型仍遵循 OpenRouter 自己的协议。
- 编程智能体捕获更干净了。Codex CLI 的托管安装可以通过 `nmem` 保存已记录的会话；Claude Code 和 Cursor 的重复捕获会稳定写入同一条对话线程。
- 本地搜索索引损坏时，搜索现在更可靠。Mem 会检测缺失的索引片段，立即提示重建，并在重建完成前保留可用的降级结果。
新增
改进
修复
- AI Now、Feed、Knowledge、Graph Intelligence 和浏览器捕获现在可以稳定运行 DeepSeek V4 工具循环，不会因为 thinking 历史缺失或不匹配而失败。
- OpenAI 兼容的自定义 DeepSeek V4 端点现在会获得和官方 DeepSeek 提供商相同的工具循环兼容保护；通过 OpenRouter 托管的 DeepSeek 模型仍保持 OpenRouter 的请求行为。
- Claude Code 会话捕获现在能在桌面文件监听和 Claude Code 插件同时开启时保持在同一条线程里。重复保存会更新同一个 `claude-code-<sessionId>` 线程，不再把一段会话拆成两条记录。
- Cursor 会话导入现在会绑定到真实对话，而不是只绑定到工作区数据库。手动导入、应用内轮询和解析器输出会收敛到同一条 Cursor 线程，不再产生重复或错误合并的记录。
- 搜索索引存储损坏现在会被当作可重建的投影问题处理，而不是显示成空搜索结果。Mem 检测到本地索引片段缺失时，会标记需要重建、刷新标题栏提示，并避免可能干扰活跃智能体写入的在线清理。
0.8.22026年5月1日
Intel Mac Search Index Hotfix
0.8.2 is only for Intel macOS desktop builds. It fixes a search-index startup and rebuild failure caused by the older Intel-compatible LanceDB package. Apple Silicon macOS, Windows, Linux, mobile, CLI, and server packages stay on 0.8.1.
亮点
- Intel Mac 用户现在可以重新初始化和重建搜索索引。Mem 已兼容 Intel 运行时里较旧的 LanceDB 表列表 API。
- 其他平台不需要这个 hotfix。如果你使用 Apple Silicon macOS、Windows、Linux、移动端、CLI 或独立 server 包，请继续使用 0.8.1。
0.8.12026年5月1日
ChatGPT MCP, Web AI Now, and Remote Fixes
0.8.1 focuses on remote-first work: ChatGPT Web and desktop can connect to Mem over Access Anywhere with OAuth, AI Now sessions live on the Mem server for desktop, web, and mobile clients, browser folder import works in Library, and several Access Anywhere, updater, Feed, and CLI edge cases are fixed.
阅读发布博客亮点
- ChatGPT 现在可以通过 Access Anywhere + OAuth 连接 Mem。复制公开 MCP URL，粘贴到 ChatGPT Connectors，在浏览器里授权后，就能在 ChatGPT 网页版或桌面版里通过 `@Nowledge Mem` 调用。
- AI Now 现在跟随 Mem 服务器。桌面端、网页端、移动端和 Access Anywhere 客户端都能看到同一份任务列表，继续同一段工作，并在 Threads 里找回已完成的 AI Now 对话。
- Library 文件夹导入现在也支持网页端。浏览器和远程客户端会保留文件夹内相对路径，并通过和桌面远程模式相同的分块导入路径上传。
- Access Anywhere 更稳了。已保存的 Cloudflare tunnel 会在启用期间持续恢复，OAuth 授权流程更清楚，已连接的云端 MCP 客户端也可以从 Integrations 里断开。
改进
- AI Now 会话现在跟随 Mem 服务器，而不是只留在某一个桌面窗口里。桌面端、网页端、移动端和 Access Anywhere 都能看到同一份任务列表，并继续打开同一个对话。
- AI Now 对话现在会在每轮完成后保存到 Threads。你在 AI Now 里的工作之后也能像 Claude Code、Codex、Cursor 等会话一样被搜索到。
- Library 文件夹导入现在支持 `/app`。浏览器和远程客户端会保留文件夹内相对路径，并使用有边界的上传 API，不再依赖仅桌面端可用的本地文件路径。
- Access Anywhere 的自定义 Cloudflare tunnel 会在启用期间持续恢复；Remote MCP over OAuth 也为 ChatGPT 和其他支持 OAuth 的客户端提供了更清楚的一条配置路径。
修复
- 当内置 Kimi 运行时比 Mem 开发树稍旧时，AI Now 回复不再消失。文本、思考过程、工具卡片和最终回复都会继续正常流式显示。
- 当连接的 Mem 服务器可以运行 AI Now 时，网页端也能直接打开 AI Now。通过 `/app` 访问 Mem 时，不再看到旧的“仅桌面端可用”占位页。
- 当浏览器标签页断开或跟不上流式事件时，AI Now 网页流不再保留过期的实时更新。长时间运行 Access Anywhere AI Now 会话时，界面会更稳定。
- 当前日期、插件说明等隐藏运行上下文不再像用户输入一样出现在 AI Now 对话记录里。
- ChatGPT OAuth 授权不再卡在同意页面；已授权的云端客户端会获得预期的 Mem 工具，但不会拿到危险的本地删除或连接器操作。
- 现在可以在「集成」里断开已授权的远程 MCP 客户端。断开后，Mem 服务器会移除该客户端和对应 token。
- Daily Briefing 里的 flag 和 insight 现在会重新打开对应的 Feed finding，不会再卡在找不到 memory 的详情卡片上。
- 更新下载完成后点击 Install，现在会明确进入重启安装流程；如果安装包丢失，或 Windows 辅助进程无法确认安装成功，也会显示更清楚的错误。
- `nmem status` 现在会根据 CLI 的安装方式提示该怎么修复版本不一致，包括桌面应用、PyPI、pipx、uv/uvx 和 WSL 场景。
0.8.02026年4月29日
Mem Becomes Your LLM Wiki
0.8.0 brings the LLM Wiki pattern fully inside Mem. Topic, entity, and crystal pages read like real wiki entries; `[[Entity Name]]` wikilinks are clickable everywhere; Study with AI turns a source into saved memories or crystals you decide to keep; Wiki Export ships your knowledge as portable markdown. Investigate carries any wiki page into the Knowledge Graph with Graph Intelligence already aimed at the right cluster. This release also adds scanned-PDF reading and closes reliability gaps in Spaces, Memories, AI Now recall, search rebuilds, provider routing, and Windows startup.
阅读发布博客亮点
- Library 现在会像 Wiki 一样打开。主题聚类把相关的结晶和实体放在一起，并显示简短摘要和提及次数，让你先看清自己的知识大致分布，再决定打开哪份资料。
- 实体、结晶和主题页面现在更像真正的 Wiki 页面。背景、相关实体、来源记忆、最近提及和小型图谱都放在同一个阅读面里，不用再跨标签把线索拼起来。
- 「Investigate」把 Wiki 和知识图谱连起来。你可以从实体、结晶或主题页进入图谱，相关节点会自动选中，Graph Intelligence 也会带着这段上下文继续分析。
- 「和 AI 一起精读」会把一份资料变成一次可控的提取过程。智能体阅读文档，与已有记忆对照，指出重合或冲突，并提出值得保存的内容。最后保留什么由你决定。
- 现在可以把 Wiki 导出成 markdown。Library 的下载按钮会生成一个 ZIP，包含首页、主题、实体和结晶页面，并使用和 Mem 内部一致的 `[[实体名]]` 链接。
- 资料库现在可以在后台读取扫描版 PDF。遇到没有文本层、像图片一样的 PDF 页面时，Mem 会只读取这些页面，并用你配置的视觉模型把文字补回搜索内容。
新增
- Wiki 导出。Library 页签栏右侧新增 Download 图标，可一键构建包含 index、主题、实体、结晶的 markdown 文件夹 ZIP，所有 `[[实体名]]` wiki 链接在 Obsidian 与任意 markdown 阅读器中都能解析跳转。
- 在 Library 任意资料上「和 AI 一起精读」。Ask 输入框上方的「精读」chip 会触发一段精心设计的提示词，让智能体仔细阅读资料、用 `memory_search` 与你已有的记忆交叉对照、指出冲突，并提出值得保存的记忆或结晶建议。智能体只提议，你来决定保留什么。
- Library 新增 Wiki 标签。按 AI 命名的主题聚类组织你的知识，让你在同一个标题下同时读到结晶和实体，不必再在三个标签间来回切换。
- 实体页现在像真正的 wiki 条目。主题背景、首次出现时间、智能体置信度、学习到的别名、提到它的记忆、相关实体，加上一张本地小型关系图，全部集中在一页。数据从图谱中实时派生，无需额外存储。
- `[[实体名]]` wiki 链接现在是结晶与生成产物中的一等公民。点击 chip 即可打开对应实体页；智能体提示词已更新，新生成的知识天然自带互链，并自动处理别名和大小写不敏感的匹配。
- 「Investigate」把 Library 页面和知识图谱连起来。实体、结晶和主题页可以打开图谱，并自动选中相关内容，让 Graph Intelligence 接着同一段上下文继续分析。
- 每一份合成都援引其源头。结晶页和生成产物页现在都会展示「由 N 条记忆合成」小节，列出滋养这次合成的源记忆。点任意一行就在原地预览那条记忆。
- 新增「按页取 wiki」的 API。三个 GET 端点会把任意一个实体、结晶或主题页渲染成 markdown，让智能体、脚本和 CLI 调用方按需取一页，不必再导出整本 wiki。
- 实体标签新增「高频实体」入口。六个信号最强的实体直接置顶，比字母排序的列表更先一步出现；一旦你开始搜索或选择类型筛选，它会自动隐藏，不会干扰你的更精确查询。
- 资料库现在会在后台读取扫描版 PDF。普通 PDF 会先完成导入，之后 Mem 只读取需要视觉模型的页面。如果缺少模型，或后台智能预算达到上限，文档会暂停并显示「继续」。
- 设置现在有独立的「图片文字」模型选择。它支持 OpenAI 兼容的视觉端点和标准 LiteLLM 路由的提供商；不支持的提供商类型会在一开始就给出清楚原因。
改进
- Claude Code、Gemini CLI 和 Copilot CLI 会在宿主提供相应生命周期钩子时保存当前会话。Transcript 捕获现在由原生集成或 `nmem t save` 完成，不再隐藏在 MCP 路径后面。
- Codex 和 Gemini CLI 包现在内置 Nowledge Mem MCP。远程 MCP 配置也更清楚：`nmem config mcp show` 会根据本机 `nmem` 客户端使用的 URL 和 API key，生成对应宿主的配置。
- 插件发布检查现在更直接覆盖主要集成路径，包括 marketplace 元数据、远程凭据处理、Claude WSL 桥接、Copilot 包边界、Hermes 重试鉴权和 camelCase hook payload。
修复
- 后台对话提炼现在会把新记忆保存在会话所属的 Space。由 Knowledge Agent 处理的长对话，不会再因为内部工具调用没有显式传入 Space，而把提炼出的记忆写到默认 Space。
- 在搜索、分页和隐藏已归档记忆时，记忆数量、浏览模式和过滤条件现在会保持一致。
- 较长的本地搜索重建，不会再在任务进行到一半时卸载 Search Index Model；较慢的机器也不会在重建中途悄悄降级。
- AI Now 在回忆旧记忆时更不容易漏掉结果；本地工具启动失败时也会说明失败原因，而不是看起来像智能体没有回应。
- Windows 安装在更新运行时时会清理旧的 bundled pandas 文件，避免旧编译依赖盖过新包而导致启动失败。
- Knowledge Graph 回填现在会显示真实进度，并停止反复重试没有可提取实体的记忆。
- 深色模式下的按钮和输入框更清楚了，包括本地模型下载按钮和文档站里的 Ask AI 输入框。
- 更长的后台智能、Feed 和 Graph Intelligence 任务现在更不容易过早触达步骤限制；该限制也可以在知识处理设置中看到。
- Access Anywhere 和局域网页端现在可以正常从 Feed 发送消息了。网页端会把消息发到你打开的 Mem 服务器，而不是误发到浏览器设备自己的 localhost。
0.7.102026年4月23日
Lower Rebuild RAM and Safer Linux Updates
0.7.10 focuses on the heavy maintenance paths. Search rebuild uses less memory, update downloads are faster, and Linux installs recover more clearly when the runtime, graphics stack, or CPU is the real blocker.
亮点
改进
平台
- Linux 安装包现在会和桌面应用一起把后端代码真正切换到新版本。DEB 或 RPM 更新后，Mem 会优先使用健康的系统打包 Python runtime，不会再意外继续卡在 `~/.local/share/nowledge-mem/runtime` 里留下的旧备用 runtime 上。
- 一些 Linux 桌面现在不需要再手动改 `.desktop` 文件，才能绕过首次启动白屏。Mem 现在不仅会在 Arch 一类系统上自动走更安全的 WebKit 启动路径，也会在检测到容易出问题的渲染栈时自动启用，例如使用 NVIDIA 驱动的 Linux，或 VMware 这类虚拟化桌面。
- 较老的 Linux `x86_64` 机器现在不会再只丢一个 `Illegal instruction` 就直接崩掉了。当前桌面安装包要求 CPU 支持 AVX2，Mem 现在会在启动前直接说明这一点，而不是在较老、没有 AVX2 的 Xeon / Ivy Bridge 一类机器上不明不白地崩溃。
0.7.92026年4月22日
AI Now Runs Smoother, Search Rebuilds Stay Stable
0.7.9 removes friction from two heavy paths: long AI Now research runs and large search rebuilds. Research stays readable and responsive, rebuilds are less likely to spike RAM or lose their place on smaller machines, and the browser extension also adds Coze and Yuanbao capture.
亮点
改进
修复
- AI Now 里的网页工具卡片和命令卡片，现在会显示真正的操作和目标，不会再出现 `Code`、`https`、`/` 这种让人发懵的占位内容。
- Research 现在不会再把页面被拦截，或站点返回的 `401/403`，误判成插件登录失效了。页面本身打不开时，错误提示会更诚实。
- 搜索重建现在更能扛住旧数据里的坏行、缺失的公开线程 ID，以及短暂的远程 embedding 故障了。一次脏数据页或一次短暂的提供商抖动，不再那么容易把整轮重建拖垮。
- Windows 和 Apple Silicon 上的搜索模型恢复现在更靠谱了，原本就有效的旧版模型布局也不会再被误判成必须重装。
- 超大的 PDF 和其他大体积资料，现在在解析和分块时更安全了；刚导入的资料也会重新立刻出现在搜索里。
- Copilot CLI 的商店安装现在不再依赖一个实际上不存在的 setup 脚本了。Mem 现在会直接使用打包好的 hook 文件，所以正常安装和更新后也能继续工作。
0.7.82026年4月21日
A Better Graph Overview, Cleaner Labels, and Easier Copilot Setup
0.7.8 makes the Graph easier to navigate. You now open into a calm overview of your whole knowledge, uploaded documents are called Artifacts throughout the graph, and Copilot CLI becomes a proper built-in integration.
新增
- Copilot CLI 现在也成了桌面应用里的一级集成入口。你可以直接在「Integrations」里看到它，判断插件是否已经安装，并让 Mem 在应用内更新插件和它的会话捕获运行时，不必再切去终端。
- 打开图谱，默认进入一个从容的 Overview（总览），一眼看见你全部的知识；想放大细看时，按一下 V 或点「Focus」标签就进入聚焦视图，用采样后的子图做精读。在总览里选中节点就会浮出一条操作条：刚好选了两个节点时，可以点 Path 画出它们之间的最短路径；按 W 把当前选择扩展到邻居；按 E 仅对选中的部分切入 Focus 视图做深入分析；按 D 清空选择重新开始。想框选某个区域时，一键 A 就能切到套索模式。
- 图谱的操作方式现在就是你已经熟悉的那种手感：点击节点打开检查面板，点击空白处取消选择；拖拽节点移动位置，拖拽空白处平移画布，任意模式下按住鼠标右键都可以平移。按 S 进入多选，按 A 拉套索，按 1–4 分别显示/隐藏 Entity、Memory、Artifact 和 Thread 四类节点。左下角的类型筛选现在是一组带颜色圆点的小胶囊，和节点调色板同一套语言，看一眼就明白，不用再去猜 E/M/S/T 这些字母代表什么。
- 图谱里现在把你上传的文档统一叫「资料」——筛选胶囊、选择摘要、对话里的迷你图都是一样的说法，和你脑子里「资料库」的分类终于对上了，中英文读起来都是一回事。
- Graph Intelligence Agent 在读你资料库里的东西时，现在会直接在画布上给你看：它每次检索、打开、搜索段落或分析表格数据，对应的资料节点都会在 Overview 里亮起来，你不用切面板、不用离开图谱，就能一边看到它在想什么、一边跟着它的推理往下走。
修复
- 共享空间里的线程，现在会按它真正所属的空间继续提炼，而不会在最后一步突然失败了。比如 Space B 被允许读取 Space A 时，你现在可以在 B 里打开 A 的线程，然后继续做提炼计划、预览和保存，不会再碰到那种其实是误判的「这个线程不属于当前空间」错误。
- Knowledge Agent 从线程里自动提取记忆时，现在终于有了真正的写入前质量闸门。Mem 会先拦住已知的占位模板文本，检查同一条线程里是不是已经生成过意思相同的记忆，再把新的候选记忆回到原始对话里做校验，同时把被夸大的重要性分数压回合理范围。这修掉了一个很伤信任的问题：一条质量不好的线程，不会再一口气往图谱里喷出模板记忆、重复记忆和被说得过重的总结。
- 长线程提炼时，也不会再在这个新的质量闸门里自己崩掉了。Mem 现在会安全地处理带 `/` 或 `-` 的 ID 和术语，在构建校验用的受限摘录时不再触发正则错误，所以越长的对话也能继续被保护，而不是偏偏在长线程上失效。
- macOS 桌面版的远程模式，现在又能正常连接纯 HTTP 的 Tailscale 地址了。如果之前同一台 Mem 服务器在浏览器里能打开，但桌面 App 只有把 `http://...` 改成 `https://...` 才能连上，这个缺口现在已经补上：桌面端内置 webview 终于和产品原本承诺的一样，允许走同样的 HTTP tailnet 路径。
- 如果 Mem 启动时恢复的是一台暂时离线的远程服务器，现在你可以直接先在这一次启动本机后端继续用，不必再在“卡在失联的远程连接上”等它恢复，和“彻底退出远程模式”之间二选一。像 Tailscale 设备睡眠、临时掉线这类短时网络问题，不会再让 App 看起来像被锁住了一样。
- 在这个恢复界面里修改已保存的远程服务器或 API 密钥时，如果重试后还是错的，Mem 现在会继续把你留在同一个恢复界面里，不会再因为一次错误重试就突然切去启动本机后端。改错字就是改错字，不会再像是被甩进了另一个模式。
- 桌面版现在不会在每次启动时，又把你的 `nmem` CLI 终点偷偷改回去了。Mem 依然会记住你上次连接的远程服务器，并按那个状态进入远程模式，但它不会再在启动时直接改写 `~/.nowledge-mem/config.json`；只有你在 App 里明确重新连接或切换模式时，才会同步。这样一来，就算保存下来的远程服务器已经失效，也不会把你刚手动修好的 CLI 配置又改坏。
- Windows 启动现在不再依赖向 `%USERPROFILE%\.cache\nowledge-graph` 写后端缓存。Mem 改为使用应用自有的 `%LOCALAPPDATA%\NowledgeGraph\cache`，在家目录缓存权限受限的机器上启动更稳定。
- `nmem` 与终端 UI 现在会对 `127.0.0.1`、`localhost` 这类回环地址绕过系统代理，不会再因为开启了全局代理就把本地请求也送错路径而失败。
- Linux 上 `nmem` 启动包装器现在对符号链接更安全。通过安装器创建的 `/usr/local/bin/nmem` 链接启动时，会正确定位到内置 Python，不会再报 “Could not find bundled Python”。
- 知识图谱的总览视图，在一些桌面环境里不再塌成一整块空白或一个小点。之前图数据其实已经加载成功，但总览画布可能会在还没拿到真实尺寸时就先执行自动适配，结果把整张图缩成了一个点。现在 Mem 会等画布尺寸有效之后，才确定第一次总览适配。
- 网页端和远程客户端现在又可以重新置顶记忆与线程了。收藏切换现在会把 `is_favorite` 标志发送到 API 真正要求的位置，所以浏览器或远程桌面客户端里点击置顶时，不会再出现验证错误，而本地桌面端却正常的分裂体验。
- 网页端 `/app` 和 Android 客户端里的「About」支持链接现在会正常打开了。GitHub、文档和支持入口又会像真正的外部链接那样工作，不会再偶尔弹出文件选择器。
- 删除资料库里的来源时，现在不会再出现“明明还在，却提示删除成功”的情况了。来源节点现在也可能带有标签关系，Mem 已改成用对关系安全的图删除方式来处理；如果删除失败，也会返回真实错误，而不是假装成功。
- Claude Code 会话保存与自动同步现在支持带点号的项目路径（例如 `jane.doe` 这类用户名或 `my.project` 这类目录名），导入流程不再因路径归一化差异而失败。
- 已归档的 AI Now 任务现在可以重新开始了。归档不再把一个保存下来的任务变成只能看的历史记录；你可以直接从归档列表里重新打开它，Mem 会在恢复任务时自动取消归档标记。
- AI Now 的网页搜索现在不会再因为没有设置地区而把整个研究流程卡住了。当你没有指定搜索地区时，Mem 现在会让搜索库使用它自己的默认值，而不是继续传一个空值进去，导致某些上游搜索引擎直接崩掉。
- 长时间运行的线程导入，现在不会再因为你点到对话框外面就直接消失了。Mem 会在解析或导入进行中锁住导入窗口，所以像 ChatGPT HTML 这种大批量导入，不会再在关键步骤中途被误关掉。
- 大型 ChatGPT HTML 批量导入现在会把消息索引刷新延后到整次任务结束后执行，不再每导入一个线程就刷新一次。这样能显著降低长任务期间的后端压力，同时保持最终搜索结果正确。
- Alma 备份 ZIP 导入现在不会再因为消息 ID 出现在时间戳相邻字段而失败。导入流程在该路径下只接受可解析的时间戳，并跳过类似消息 ID 的值，不会再逐线程崩掉。
- Codex 的安装指引现在改成了更符合真实体验的路径：先添加 marketplace，再安装 `nowledge-mem@nowledge-community`，在 `~/.codex/config.toml` 里启用插件，并把 Nowledge Mem MCP 一起配上，让 Codex 不再只停在 Working Memory，而会更主动地继续检索。
- Copilot CLI 的更新步骤，现在按 GitHub Copilot 真实要求的方式来写了。Mem 会明确告诉你先刷新 marketplace，再运行 `copilot plugin update nowledge-mem`，并同步更新捕获运行时，而不是继续显示那个容易和宿主漂移的旧版带 marketplace 后缀命令。
0.7.72026年4月20日
Library Search Reaches More Places
0.7.7 brings Library search to more AI surfaces and tightens a broad set of reliability issues. AI Now, Feed, Graph Intelligence, and `nmem` can all reach into your documents, while search recovery, updater behavior, and provider diagnostics all get less fragile.
亮点
- 你上传的文档，现在每个 AI 工作面都能用。AI Now、Feed Agent、Graph Intelligence Agent 和 `nmem` CLI 都可以搜索、阅读并分析你资料库里的内容，不再只有 Knowledge Agent 能用。
- `nmem sources` 新增 `search`、`read`、`search-chunks` 和 `analyze` 四个子命令，终端与脚本工作流现在可以直接检索资料库，不必再打开应用。
- AI Now failure reports now include the real provider error text from `kimi.log`, so copied diagnostics actually explain what went wrong instead of only showing the last harmless lines.
新增
- 资料库里的文档现在可以在 AI Now、内置 Feed Agent 和 Graph Intelligence Agent 里一起搜索，不再只限于 Knowledge Agent。向其中任一位 Agent 询问一份你已上传的 PDF、电子表格或 Word 文档，它们现在会真正去读取已解析的内容并引用匹配段落，而不是只能根据一些恰好提到该文件的记忆摘要来回答。
- `nmem sources` 新增 `search`、`read`、`search-chunks` 和 `analyze` 四个子命令。你可以直接在终端里按文件名或摘要检索资料库、按偏移量读取解析后的内容、在单个文档内部搜索匹配段落，或对 CSV、电子表格进行表格分析。
- 通过 MCP 连接 Mem 的外部 AI 工具，继续保持一个有意更聚焦的资料库工具集（检索、读取、片段搜索与分析）；像删除这种管理操作则保留在桌面应用和内置 Agent 中。这一策略现在已经在代码里写明，后续审查不会再漂移。
修复
- 在单份资料库文档内部搜索现在会返回真正的匹配文本、片段位置和分数，而不再是空字符串和 `null` 位置。`nmem sources search-chunks` 的输出以及等价的 Agent 工具现在都包含真实的匹配段落，你和 Agent 都能直接看到匹配到的内容，而不只是匹配到了多少个片段。
- 针对内置 Agent 的提示词与工具列表的修改，不会再看起来只生效一次、之后又被悄悄覆盖。早先放在代码旁边的 Agent 文件，会在每次会话启动时被打包内的默认值重写；几个月里不同 Agent 之间对自己能用哪些工具出现分歧，原因就在这里。Mem 现在对 Knowledge、Feed 和 Graph Intelligence Agent 的提示词只有一份唯一源：一次修好，就在所有地方保持生效。
- 浅色外观下创建新空间时，输入内容现在重新变得可见。共享输入框已经改为使用主题自己的正文色和占位色，所以你输入的文字不会再只有选中时才看得见。
- `nmem` CLI 和终端界面现在在访问本机 `127.0.0.1` / `localhost` 的 Mem 时，会主动忽略系统代理设置。之前在 Windows 上，只要机器开了全局代理，本机命令就可能失败，哪怕桌面 App 本身是正常的。现在这类 loopback 请求会直接走本机，不会再被错误地拖进代理链路。
- Linux 安装里把 `nmem` 放进 `/usr/local/bin` 之后，现在也能按正常的软链接方式启动了。生成出来的包装脚本会先解析它真正所在的安装目录，再去查找内置 Python，所以不会再因为你是从安装器创建的软链接启动，就一上来报“Could not find bundled Python”。
- 连接到远程 Mem 服务器的桌面客户端，现在重新可以顺利退出了。即使你的数据连接指向另一台机器，关闭界面也不会再卡在最后的“正在安全关闭...”这一步。
- AI Now 里的记忆搜索卡片现在会显示真正返回的记忆结果。之前搜索明明成功，却仍可能显示为空，是因为卡片只认一套整理过的结果结构，没有读取 MCP 原始响应里的 `memory_id` 和 `total_found` 这些字段。
- AI Now 的失败诊断现在会复制 `kimi.log` 里真正的报错摘录，而不再只带上最后几行看起来正常的日志尾部。如果是提供商拒绝了请求，复制出来的诊断信息更容易直接带上真实异常，而不是只剩下 `Connected MCP server: nowledge-mem` 之类的行。
- ChatGPT Subscription 现在只会在设置里提供兼容 Codex 的模型选择，因此像 `gpt-5-chat-latest` 这样的 GPT-5 聊天别名不会再被保存进 AI Now 然后在第一轮就失败。即使你之前已经保存了不兼容的旧选择，Mem 现在也会直接告诉你切回 Codex 模型，而不是只给一个模糊的提供商错误。
- 当旧的 LanceDB 搜索投影已经损坏时，搜索重建现在会先从一组全新的派生搜索表重新开始，而不是试图在原地清空并复用损坏表状态。这样修复了那类“看起来重建完成了，但记忆向量搜索仍因旧投影引用缺失文件而退回文本搜索”的恢复问题。
- 如果搜索重建已经在运行，Mem 现在会明确拒绝第二次重建请求，而不是让两个重建流程同时争抢同一份搜索索引文件。
- Morning Briefing 的重试与去重现在更可靠：定时任务失败不会再阻断当天手动重试，排队/运行中的触发不会再并发出重复任务，处理状态也会区分“最近一次尝试”和“最近一次成功完成”。
- 远程与 Web 客户端现在可以正常验证并应用 embedding 设置，不会再被 remote mode 的 403 网关错误拦截；远程 embedding 模型检测与“是否需要重建”判断重新可用。
- Windows 应用内更新现在会从稳定的应用根目录安装并重启，已下载更新在安装前也会保留自己的版本身份，减少“看起来已更新但重开还是旧版本”的失败场景。
- macOS 构建现在可以更稳定地访问局域网里的 AI 端点（包括 homelab 的 OpenAI 兼容服务）；若仍被本地网络权限阻断，提示信息也更清晰。
0.7.62026年4月17日
Search Recovery and IME Input Hotfix
0.7.6 fixes two problems that show up fast in real use: search rebuild recovery on upgraded indexes, and accidental message sends while typing with CJK IME input.
亮点
- Search metadata upgrades now run in bounded exact-row batches and stay non-destructive, so large upgraded indexes no longer get stuck in prolonged metadata-update loops.
- Full search rebuild now pages graph data and streams thread messages in bounded batches, reducing peak memory pressure and recovering stale metadata-update states from older marker shapes.
- AI Now and Graph Agent inputs now respect CJK IME composition, so Enter confirms input without accidentally sending unfinished messages.
修复
- AI Now 与 Graph Agent 输入框现在会正确处理中日韩 IME 组合输入：按 Enter 时先确认输入，不会再误发半成品消息。
- 内置 Feed Agent 与 Knowledge Agent 的线程搜索现在走同一条空间作用域搜索路径，后续修复会同时覆盖摘要片段、来源过滤与空间结果格式，不再在两份实现之间漂移。
- Memory 與 Thread 的批次移动/删除现在共用同一套后端工具，用于稳定去重 ID、解析来源/目标 space，并约束同一 space 内的大批量选择。
- 搜索元数据升级不再按 thread 或 source 逐个重写 LanceDB。0.7.6 改为按精确行分批回填，优先使用已存储的 thread 身份，并且只有在确实存在更新标记时才启动后台修复。
- 完整搜索重建不再试图一次把过多图谱数据堆在 Python 内存里。0.7.6 会分页处理 memories、sources、entities、communities，按有界批次流式写入 thread messages，并且能恢复那些因旧版标记形态而卡在元数据更新状态的安装。
- 仅元数据的搜索升级现在改为用带合并语义的方式回填 LanceDB 行，而不是先删后加。这样即使元数据回填中途被打断，也不再容易把刚修复好的搜索行暂时从派生索引里删掉。
- 完整重建成功后，现在也会一并清理过期的元数据更新状态；对于非常大的 Library 文件，也不再需要在开始 chunk 索引前先额外保留一整份切块列表在内存里。
0.7.52026年4月16日
Bulk Space Cleanup and Export Reliability
0.7.5 makes large space cleanups more practical. You can move or clean up Memories and Threads in bulk, export is steadier on busy setups, and OpenAI-compatible providers plus localhost browser-extension access both behave more predictably.
亮点
- Memories and Threads can now be moved or cleaned up across pages in one operation, with final selection resolved safely on the backend.
- `nmem m move` and `nmem t move` now support named-space migration from terminal workflows, including dry-run previews for whole-space changes.
- Data export now finishes reliably even when Ladybug drops the read-only transaction before `COMMIT`; Mem preserves the archive and records a warning instead of failing the download.
- OpenAI-compatible providers are now validated and streamed more safely, including `/responses` endpoints that return SSE and backends that do not support optional usage-stream extensions.
- Browser extension and secure AI sites can again reach local Mem on `127.0.0.1` without forcing Access Anywhere, thanks to explicit Private Network Access preflight handling.
新增
- Selection mode now supports visible-page or full-current-space Memory/Thread cleanup and move operations without pushing massive client-side ID lists.
- Terminal-first migration is now supported with `nmem m move` and `nmem t move`, including dry-run previews for whole-space moves and TUI lane-aware browsing via `NMEM_SPACE`.
修复
- Unix 启动与后端清理流程现在更稳：启动前清理已移出 UI 线程并为子进程设置超时，停止旧后端时也沿用同样的超时约束，不再因系统工具卡住而导致启动/重启挂起。
- Opening Memories no longer competes with hidden page preloads. Desktop now defers non-visible view fetches and avoids duplicate first-page requests while parent queries are still loading.
- Data export is more reliable on busy or remote setups: read-only `COMMIT` races are treated as warnings in the archive, and export now fails explicitly if a dedicated export connection cannot be opened.
- Timeline Knowledge Pulse now shows crystals and communities correctly in active-space views (not just global view).
- Remote and web model pickers now keep the selected provider when listing models before save, preventing wrong-provider model lists in temporary edit flows.
- OpenAI-compatible `/responses` providers that return SSE for validation can now be verified correctly instead of being rejected for non-JSON validation replies.
- The browser extension and secure AI sites can again reach local Mem on `127.0.0.1` without forcing Access Anywhere, thanks to explicit Private Network Access preflight handling.
- Space-scoped thread actions now work on older databases with duplicate preserved thread IDs across spaces by honoring explicit space guards end-to-end.
- Thread search now stays inside the selected space and avoids borrowing globally ranked hits from another lane when an older index needs rebuild.
- Scoped thread and Library search now keep space metadata in the search index, and upgraded indexes repair metadata-only gaps automatically without a full semantic rebuild.
- Shared-space thread results now open correctly from the same space where you found them.
- Thread identity migration and thread search now tolerate older duplicate conversation rows safely: startup backfill no longer fails on raw Ladybug result iteration, and ambiguous legacy all-space hits are skipped instead of hiding valid scoped results.
- Exports that exclude threads now stay truly thread-free: `ThreadIdentity` and thread-linked relations are filtered to avoid dangling references.
- OpenAI-compatible streaming no longer assumes optional `stream_options.include_usage` support, avoiding unnecessary failures on strict compatible backends.
- Background Feed Agent no longer crashes on non-standard OpenAI-compatible response shapes; errors are now handled and surfaced clearly.
- Copilot and Codex token failures now report actionable cause (`inactive subscription` vs `revoked token`) instead of generic auth-expired messaging.
0.7.42026年4月14日
Search, Thread Sync, and Remote Clients Feel More Trustworthy
0.7.4 is a reliability release. Search explanations stay more honest, thread sync causes less UI drag, and desktop, web, remote, and mobile clients stay closer to the same state.
亮点
- Pinned threads now behave as one thread everywhere: pin/unpin works reliably without duplicate rows or key-warning side effects.
- Background session sync no longer makes Mem feel frozen while Claude Code/Codex/Cursor/OpenCode imports catch up.
- Search explanations and per-result score factors now stay honest and query-bound, with clearer product-language breakdowns for Meaning/Text/Keywords/graph hints.
- Search rebuild now validates real semantic indexing readiness and no longer marks a text-only fallback as a successful semantic rebuild.
- Remote/mobile space switching and Settings usage counts now reflect real server state instead of stale fallback data.
- Graph Memory Auto now stays truly automatic after OOM recovery instead of silently becoming a stale fixed limit.
改进
- Claude Code integration updates now follow Claude's official upgrade path instead of marketplace-guess heuristics, reducing plugin drift during updates.
- AI Now now works cleanly when loopback API auth is required, so local-only secured setups no longer break chat startup.
- Linux 上的 `nmem update` 现在会区分“版本已发布”和“APT 已经能安装这个版本”，引导命令也改成了 `install --only-upgrade`，不再使用误导性的定向 `upgrade` 写法。
- Quitting the desktop app now waits for final save handoff before backend teardown, reducing last-moment data-loss races.
- Streaming OpenAI-compatible generation no longer blocks the rest of backend API work while one reply is still arriving.
- Background watcher sync now parses large transcript files off the API loop and serializes watcher-triggered imports, so normal Mem API surfaces stay responsive while sync catches up.
- Threads now keep canonical public IDs across read/delete/append/reconcile paths, so cross-space sync/import flows no longer clone the same conversation under drifting identities.
- Step-limited thread analysis now pauses cleanly without pretending a thread was fully analyzed or immediately hot-looping the same work.
- Remote thread distill and knowledge-graph preview paths now return on bounded interactive contracts instead of paying background-grade polish cost before preview.
- Knowledge graph extraction now follows your configured remote text model by default instead of silently drifting to local MLX.
- Search details now consistently use product language and preserve explicit per-result factors, so ranking explanations remain understandable and numerically grounded.
- Fast search quality is improved for broad multi-word queries with better cheap keyword extraction and full-query coverage weighting.
- Deep search and reranking metadata now stay robust under edge-case result shapes, so strong candidates outside the first cut can still surface correctly.
- Deep-search explanations now stay tied to the request that produced them even after later post-processing stages, preventing cross-query explanation drift.
- Fast search keeps primary meaning/text evidence stronger and avoids diluting relevance with non-retrieval bonus noise.
- Graph search quality is improved for focus-mode queries: weaker tail matches are trimmed, memory seed ranking is aligned with trusted memory relevance, and layout settling is clearer.
- OpenRouter attribution now consistently identifies Mem (desktop agents, background jobs, browser extension) instead of mixed SDK/app labels.
修复
- Pinned threads no longer duplicate or refuse to unpin due mixed ID/favorite state paths; pinned sidebar and thread view now stay in sync.
- Thread unpin now updates immediately in open detail view and pin-state toggles stay visually consistent right after fresh toggles.
- Pinned memories and threads no longer disappear when early favorites windows include duplicate or false rows; favorites views now stay stable.
- Remote/mobile clients now switch spaces reliably without stale fallback-roster revalidation blocking the selected lane.
- Settings space usage counts on remote/web now come from the correct full snapshot path instead of incorrectly showing all spaces as empty.
- Search rebuild now validates semantic backend readiness and reports underlying startup dependency errors instead of generic failure labels.
- Local semantic search gets a longer first-query warmup budget and no longer falls back too early while the embedding service is still loading.
- Imported coding-agent scaffold/system prompts are excluded from distill/extract relevance views and provenance so extracted memories stay focused.
- Session append fast path now stays fast and append responses keep valid message counts even when sparse internal rows are returned.
- Turning off automatic thread analysis now truly disables that path instead of continuing background analysis unexpectedly.
- Threads browse pagination now stays page-sized even with historical duplicate rows, and legacy blank-space thread rows no longer appear twice in Default.
- Thread analysis state and pinned-thread truth now stay scoped correctly across spaces and repeat sync cycles, avoiding cross-space state drift.
- Memory labels now behave consistently across clients and views: creation/removal paths use canonical contracts and no longer report fake success on no-op failures.
- AI Now no longer black-screens on a single rich-reply renderer failure; rendering errors are isolated to the affected message.
- AI Now input reliability is improved: first-question composer clearing and resume-queue handling no longer leave stale or misflushed messages.
- Feed URL capture now falls back to rendered-page text for low-signal SPA shells (for example X), preserving real post content instead of shell markup.
- Data export now uses a dedicated read connection to avoid one transaction race that could surface as `No active transaction for COMMIT` on busy apps.
- Library export now includes the source records and files users expect, not only graph rows.
- Browser-extension Distill now surfaces the real failing boundary instead of a bare `Failed to fetch`, making troubleshooting actionable.
- Graph Intelligence chat now works in remote/web mode again, no longer failing with misleading auth errors; it also handles more OpenAI-compatible response-shape edge cases safely.
- Graph Intelligence chat no longer drops mid-conversation when GitHub Copilot or Codex tokens expire; token refresh now keeps sessions alive.
- Graph controls text is now readable in light mode.
平台
- Graph Memory Auto now preserves auto mode after OOM recovery with explicit floor/cap behavior and reset controls, instead of silently degrading to stale fixed limits.
- macOS non-admin CLI installation now completes as a usable terminal command by updating the active shell startup path and cleaning stale fallback PATH blocks when no longer needed.
0.7.32026年4月12日
Follow-up Hotfix
0.7.22026年4月12日
Spaces: Timeline & Feed Hotfix
0.7.2 fixes a regression from Spaces where Timeline and the AI Feed could use the wrong lane or show empty previews. Your feed and timeline again follow the active space consistently with the rest of Mem.
亮点
0.7.02026年4月12日
Introducing Spaces
0.7.0 introduces Spaces, the first real way to separate projects, purposes, and agent tendencies inside Mem without forcing everything into one shared pool. The same release also brings a broad performance pass, so larger knowledge bases feel faster.
亮点
新增
- Introduced Spaces as Mem's optional lane model for project and purpose isolation. You can keep focused lanes for different agents or workflows while keeping `Default` as the simple single-lane path.
- Added shared space profile APIs and CLI coverage so spaces are managed as a product-level contract, not as one-device local state.
- AI surfaces and background workflows can now inherit the active space context by default, including per-space Working Memory behavior.
改进
- Space profile settings and docs now explain lane behavior in user language: retrieval scope, shared-context links, and guidance that shapes agent behavior.
- Read-path performance was optimized across memories, graph/search, and timeline-adjacent flows to reduce startup and interaction latency on larger datasets.
0.6.192026年4月8日
Graph Reasoning, Simpler Remote Setup, and Better Recovery
0.6.19 turns the graph into a place to work, not just look. The new Reason panel lets an AI agent explore your graph step by step on canvas, while remote setup, Linux headless deployment, Windows updates, and CLI text handling all get sturdier.
亮点
- Graph view now includes a Reason panel: an AI agent can explore your knowledge graph with multi-step tool chaining, while each step is visualized on canvas and can be replayed.
- Remote client setup is simpler: `nmem config client show|set|clear` gives one direct CLI path to set URL/API key for this machine and reuse it across integrations.
- `nmem t save --from opencode` is now supported in standalone CLI, including both SQLite and legacy JSON OpenCode session layouts.
- Linux headless deployment is much more dependable end-to-end: service install, port handling, local web login discovery, and upgrade recovery now avoid common first-run traps.
- Windows reliability improves in two critical paths: updater now verifies the installed target version before relaunch, and CJK/non-ASCII text no longer garbles in `nmem` output.
新增
- Graph Intelligence Agent is now built into Graph view: ask questions in the Reason panel, get multi-step graph analysis, and replay each canvas action (highlight, path, community, rank focus) directly from the conversation timeline.
- Graph canvas now includes neural visual effects during agent exploration: ripple rings, breathing glow halos, sparkles on path nodes, and softer pulse behavior for multi-selection.
- Standalone CLI now supports `nmem t save --from opencode`, including both new SQLite sessions and legacy JSON session layouts.
改进
- Remote setup is clearer and less error-prone across terminal and plugin workflows. `nmem config client` now clearly separates per-machine client connection settings from server-side Access Anywhere exposure settings.
- Gemini embedding retries are now endpoint-aware, so Mem can recover from route mismatches automatically and keep indexing/provider checks working without model-name guesswork.
- Background knowledge tasks now consistently follow your profile language and custom instructions, and morning brief handling is less likely to be suppressed by contract drift.
- Agent Sessions import/sync behavior is clearer: import saves immediately, Auto-Sync appends later messages, and the docs/UI now explain this contract explicitly.
修复
- Graph Reason panel mini-graph now preserves labels and edges for agent-discovered nodes, instead of showing truncated IDs or disconnected paths when nodes are outside current overview data.
- When graph memory is auto-raised for next launch, the app now surfaces a clear restart-needed signal and shows current vs next-launch memory values in Database Tuning.
平台
- Linux headless workflows are hardened across setup, service install, and local web login: safer service ownership/port handling, clearer startup hints, better upgrade recovery, and no false green install state when service startup fails.
- Windows in-app updater now verifies that the installed app version actually advanced to the target release before relaunch, and preserves installer artifacts on failure for easier recovery.
- CJK and other non-ASCII text no longer garbles in Windows CLI output: UTF-8 bootstrap now runs before console output, and desktop-side process env sets `PYTHONUTF8=1`.
0.6.182026年4月7日
A Faster Graph, Vision in AI Now, and Plugin Fixes
0.6.18 makes the graph much easier to explore, lets AI Now accept pasted images, and cleans up reliability across Alma, OpenClaw, Hermes, Codex, and remote settings sync.
亮点
- Graph exploration is redesigned: Overview mode renders your entire knowledge graph on Canvas 2D at 60 FPS with Navigate, Select (S), and Lasso (A) interaction modes. Select two nodes and press Q to find the shortest path between them.
- AI Now accepts pasted clipboard images directly in the chat box. If your model supports vision, the image is sent as a real multimodal prompt part.
- Alma plugin no longer freezes during use. All data operations are now async HTTP instead of blocking CLI calls, fixing multi-second UI hangs.
- OpenClaw dreaming integration: your Nowledge Mem knowledge graph is now searchable through OpenClaw's recall pipeline and participates in light, deep, and REM dreaming cycles.
新增
- Graph: redesigned dual-mode interaction. Overview shows all nodes with Navigate/Select/Lasso modes, node type filter pills (Entity/Memory/Source/Thread), floating action bar with Path (Q), Expand (W), Community, and Focus (E) actions. Left-hand keyboard shortcuts optimized for mouse-in-right-hand workflows.
- Graph: animated marching-dash effect on newly expanded edges so you can see exactly which connections were just added
- Graph: shortest path between any two selected nodes, highlighted with a pulsing golden glow
- AI Now supports pasted clipboard images for vision-capable models. Copy an image, paste into the chat box, and it is sent as a real multimodal prompt attachment.
- OpenClaw dreaming integration (community plugin v0.8.0+): your entire knowledge graph becomes searchable through OpenClaw's recall pipeline and participates in dreaming cycles
修复
- Alma plugin no longer freezes the host app. All data operations use async HTTP instead of blocking CLI calls that hung the event loop for seconds. Also fixes thread duplication on restart and broken search/recall/dedup from parameter mismatches.
- 远程连接 Mem 时，Token 预算与后台智能相关设置会与服务器保持同步：界面即真相，保存的修改会真正作用到托管知识库的那台设备。
- Browser-extension and imported ChatGPT threads now share the same identity path, so extension-synced and imported versions of the same conversation no longer split into duplicates
- Graph importance no longer drops to 0% after PageRank recalculation. Importance now always reflects the original knowledge importance; PageRank is used only for layout sizing.
- Graph shortest path query works correctly with KuzuDB (previously failed due to parameter binding in path bounds)
- Background intelligence tasks no longer get stuck in retry loops when the agent sends empty search queries
- AI Now clipboard image paste survives real desktop sends and large screenshots no longer overflow the transport connection
- Codex plugin: YAML skill descriptions no longer break serde_yaml parsing, and skill naming adjusted for Codex compatibility
- Hermes plugin: fixed install path (was writing to wrong directory) and Python import errors under Hermes' importlib loader
- OpenClaw plugin: Working Memory now injects on short messages (previously skipped on messages under 3 characters), and scheduled cron runs no longer crowd your Threads list
- 时间线与后台任务现已正确遵循您的首选输出语言设置。每日简报、洞察和知识结晶将使用您在个人资料中选择的语言生成。
0.6.172026年4月5日
Hermes Abstraction, Safer Upsert, and Better Codex Setup
0.6.17 is mainly reliability work. Hermes gets a cleaner memory-plugin path, Codex installs behave better, and upsert-by-ID is safer across CLI and integration workflows.
亮点
- Hermes plugin now follows a native memory-provider abstraction, making cross-tool memory flow more predictable than older prompt-only setups
- Codex plugin install paths and version detection are fixed, so plugin setup and update awareness are much more reliable
- Memory upsert-by-ID is now first-class: create-or-update behavior works across API and Hermes save flow to reduce duplicate memories
- Long conversations are easier to query: thread-level message search is now available across agent/tooling surfaces, and ranking is more authority-aware
新增
- Hermes memory-provider plugin abstraction is now the recommended path: lifecycle-driven Working Memory load, recall, and save behavior
- Memory upsert-by-ID support is available in core save paths for deterministic create-or-update workflows
- Hermes `nmem_save` now supports `--id`, enabling stable update-in-place memory writes from recurring workflows
- Within-thread message search is now wired across Feed tools, knowledge-agent flows, and MCP thread tooling for large conversation handling
改进
- Hermes setup guidance now clearly separates plugin-mode behavior from MCP fallback behavior to avoid stale setup patterns
- Desktop plugin management now probes Codex plugin versions for better update awareness
- `memory_add` idempotency hinting is corrected for MCP clients, improving autonomous save behavior when IDs are used
- Memory ranking is now authority-aware: stronger behavioral confidence signals contribute more directly to retrieval order
修复
- Codex plugin install paths/config/manifest are aligned with the plugin architecture, fixing broken install/update flows
- Codex plugin version metadata is restored so desktop update detection works correctly
- Memory upsert now updates `source` correctly instead of preserving stale source metadata on update path
- Graph importance analysis now preserves memory PageRank classification correctly instead of dropping importance output on save
- Background agent now rejects stale OAuth token artifacts in generated config, reducing recurring unattended 401 auth failures
0.6.162026年4月3日
OpenCode Support and Lower RAM Pressure
0.6.16 adds the OpenCode plugin and the new Exchange browser extension release, while cutting idle RAM use and making search and database memory behavior much steadier.
亮点
- OpenCode plugin is now first-class: eight tools for Working Memory, search/save, thread recall, session capture, handoff, and status
- Nowledge Mem Exchange (Chrome extension) v2.0.113 is live: saved memories are cleaner on reasoning models and capture validation is more stable
- MCP clients now receive proactive usage guidance by default, so agents are much more likely to read Working Memory and save durable decisions automatically
- AI Now, Feed preview, graph sample, and thread detail reads are much less likely to stall when background processing is active
- Idle memory on Apple Silicon drops from 8+ GB to under 800 MB over long runs, with cleanup that prevents multi-GB drift
- Update prompts now include direct manual-download links for the correct platform installer if in-app update cannot proceed
- Updater behavior is calmer: downloads start only when you choose, and status now clearly distinguishes "not ready for this platform" from "already up to date"
新增
- OpenCode integration: native plugin with eight tools (Working Memory, search, save, update, thread search, full session capture, save handoff, status) and built-in behavioral guidance
- Nowledge Mem Exchange (Chrome extension) v2.0.113: reasoning-model artifact cleanup plus stricter memory schema limits for cleaner capture and fewer failed writes
改进
- MCP server instructions now include proactive behavior guidance, so connected agents know to read Working Memory early, search when needed, and save durable decisions
- `memory_add` tool description is now explicitly proactive, increasing autonomous save behavior across MCP clients
- Hermes setup docs now treat behavioral guidance as required, not optional, which prevents the common “recalls but never saves” failure mode
- Update prompts in title bar and Settings now include direct manual download links matched to your platform as a clean fallback path
- Desktop updates no longer auto-start downloads as soon as a release is detected; download is now an explicit user action and specific versions can be skipped from the update bubble
- Updater status is now clearer when your platform package is still rolling out or when in-app updater fails, with direct recovery path kept visible
修复
- Idle memory drops from 8+ GB to under 800 MB on Apple Silicon after extended use, by cleaning up temporary model memory and reducing background memory accumulation
- Graph search no longer fails with "buffer pool is full" on larger databases; the buffer pool now scales automatically with database size, with manual override in Settings
- Exchange extension v2.0.113 now strips leaked reasoning blocks (`<think>`) and stop-token artifacts from saved memory content and capture UI output
- Exchange extension v2.0.113 now enforces memory title/content limits in tool schema (200 / 8,000) to match backend constraints and reduce validation failures
- Loopback auth no longer races startup or poisons local desktop requests after key rotation; local key bootstrap and one-time refresh logic are now reliable
- Windows search-model installs now report native ONNX Runtime load failures honestly instead of misleading "fastembed is required" package-missing errors
- Windows in-app updates no longer loop on re-downloading the same release after restart; installer outcome handling is now explicit and safer
- Platform update badges now wait until your platform’s installer is actually available, preventing dead-end update prompts during partial rollouts
- AI Now and graph-heavy reads are less likely to stall when background knowledge work is active, because remaining request-path reads now stay off the API event loop
- Feed live preview and `/graph/sample` now use bounded 1-hop graph reads designed for preview, reducing buffer pressure and freeze-like behavior during typing/exploration
- Opening a thread is read-only again: hidden self-heal writes/checkpoints were removed from GET path, reducing long pauses in thread detail under load
- Memory importance no longer drops to zero after graph importance analysis; memory PageRank is now classified and saved correctly again
- Memory search now respects your saved importance more directly, and graph-central memories can rise earlier when the text match is similarly strong
0.6.152026年4月1日
Remote Embeddings and a Friendlier CLI
0.6.15 expands remote-first workflows. Search indexing can use remote embedding providers, the CLI becomes easier to automate, and AI Now, graph, search, and background jobs all get a broad stability pass.
亮点
- Pi integration: five composable skills for Working Memory, recall, and distillation, installed with one command
- Hermes Agent integration: your cross-tool knowledge graph is now available in Hermes via MCP, with full tool access and graph exploration
- OpenCode integration: native plugin with eight tools including Working Memory, search, save, session capture, and handoff, installed via OpenCode's plugin system
- Crystal review workflow: confirm or dismiss agent-synthesized crystals, with ranking impact and clear review state in cards
- Speaker attribution in distillation now separates explicit user decisions from AI suggestions to reduce false certainty
- Bulk import now supports Claude export ZIP and Raycast AI exporter JSON in both UI and `nmem t import --file`
- Alma full-library backup ZIPs now import with thread source `alma`, matching the native plugin so filters and provenance stay consistent
- Stabilization pass for 0.6.14 remote embeddings: status, download progress, install checks, and reindex behavior now stay consistent across remote and local index paths
- Quit/update now use a graceful checkpoint-first backend shutdown path, reducing WAL leftovers and shutdown-time data-loss risk
新增
- Crystal review status is now first-class for agent-synthesized crystals: unreviewed, confirmed, and dismissed states with explicit controls and ranking effects
- Distillation now captures speaker attribution across local MLX, remote LLM, and browser-extension paths, distinguishing confirmed user decisions from AI recommendations
- Claude account export ZIP (from Claude data export) is now supported in Bulk Import and `nmem t import --file`; selected threads import first, then profile memory can be saved as labeled memories
- Raycast AI exporter JSON is now supported in Bulk Import and `nmem t import --file` for multi-session archive import
- Pi integration: five skills (Working Memory, search, distill, save-thread, status) packaged for one-command install with `pi install npm:nowledge-mem-pi`
- Hermes Agent integration: Nowledge Mem available as an MCP server in Hermes with full tool access, graph exploration, and behavioral guidance for global or project-level use
- OpenCode integration: native plugin with eight tools (Working Memory, search, save, update, thread search, session capture, handoff, status) registered via OpenCode's plugin system
改进
- Remote embedding providers introduced in 0.6.14 are now in a much more stable shape for daily use, with cleaner state handling between remote and local search-index model paths
- Threads → Import guidance is now leaner and easier to keep accurate: one canonical guide link with compact format-specific footnotes
- Import onboarding docs (EN/ZH) and Integrations listings distinguish Alma live sync vs ZIP backup, and Raycast launcher vs Raycast AI exporter
- Alma and Raycast guides (English and Chinese) now link directly to bulk import; the Chinese Bub page keeps related links in Chinese
- Mem website integration docs (EN/ZH) now separate live integrations from archive import paths more clearly for Alma and Raycast
- Integrations browser-extension copy clarifies focused-session capture versus full vendor archives
修复
- Alma backup bulk import uses thread source `alma` to match the native Alma integration
- Apple Silicon no longer silently falls back to local MLX when remote distillation paths fail; failures now surface clearly with actionable guidance
- Remote search indexing now correctly treats configured embedding providers as valid backend state across Processing status, reindex paths, and startup checks
- Custom embedding providers now use the correct OpenAI-compatible embedding path, including safer model-ID normalization
- Search Index Model install verification is more reliable on Apple Silicon across different Hugging Face cache layouts
- Search Index Model downloads no longer get stuck at "Starting download..." after successful load or duplicate-start drift
- Local Search Index Model download now works correctly even when a remote embedding provider is configured
- Local Search Index Model card now shows real progress under remote-embedding mode instead of hiding in-progress state
- Title-bar Search Index Model hint now follows real backend state and no longer sticks in false in-progress states
- Switching to remote embeddings with different vector dimensions now rebuilds index tables correctly, avoiding reindex cast failures
- Remote embedding calls no longer flood logs with LiteLLM debug chatter; normal operation is quieter while real errors still surface
- Desktop quit and in-app update paths now use unified graceful checkpoint-first backend shutdown, reducing shutdown-time storage risk
0.6.142026年3月31日
Remote Embeddings, Agent-Friendly CLI & Reliability Upgrade
This release expands remote-first workflows and tightens stability across AI Now, graph, search, and background processing. You can now run search indexing with remote embedding providers (Pro), use a much more automation-friendly `nmem` CLI, and rely on calmer AI Now sessions with stronger retrieval and rendering behavior.
亮点
- Remote embedding models (Pro): use OpenAI/Voyage/Cohere-style embedding APIs for search indexing without downloading local embedding models
- `nmem` is now more agent-friendly: stdin input, dry-run deletes, richer success output, and safer non-interactive behavior for automation
- Codex CLI plugin is now first-class in Integrations install/update flow, with one-click management alongside Claude Code, OpenClaw, and Bub
- AI Now now pulls from threads and Library more reliably, renders retrieval cards more clearly, and stays calmer during heavy tool output
- Web and remote Library uploads are more stable for large files, and remote folder import summaries now behave more like desktop imports
- Output Language and Custom Instructions now apply more consistently across Feed, extension distill, and knowledge extraction
- Graph and scheduler reliability improved: cleaner neighbor expansion, readable light-mode labels, and token-budget pause state now survives restart
新增
- Remote embedding providers (Pro): configure embedding purpose on remote providers and auto-detect output dimensions with compatibility checks and rebuild prompts
- Agent-friendly CLI upgrades: stdin content support, `--dry-run` delete previews, better `--help` examples, non-TTY-safe behavior, and richer write confirmations
- Integrations one-click plugin management now explicitly supports Codex CLI plugin install/update workflows in-app
修复
- Knowledge Graph neighbor expansion is now stable and readable: expanded nodes are laid out around the selected node with cleaner link handling
- Graph labels stay readable in light mode and stale analytics warnings are reduced in console output
- Graph hover cards no longer mask node clicks: detail cards are offset with a short delay so inspection doesn’t block selection
- Sidebar utility actions stay visible under display zoom and smaller effective heights, so bottom actions no longer collapse away
- Display size preference now uses native desktop zoom for better click accuracy and full-window rendering at smaller scales
- Knowledge Agent token-budget pause now persists across app restart, so hourly/daily enforcement and Settings state stay accurate
- Headless remote-access installs now follow the same stable CLI config path contract as desktop installs
- Search provider setup is calmer and more context-aware: startup setup no longer pushes local model download when remote search is already configured
- Search no longer breaks on older memories with null analytics fields: nullable metadata is normalized before result hydration
- Morning summary no longer fires several times a day: a 30-minute cooldown prevents the daily briefing from cascading into follow-up refreshes, and job registration deduplicates on restart
- Labels now appear on the memory card immediately after you add them, without needing to navigate away and back
- Knowledge Agent no longer retries invalid tool calls endlessly; per-signature and total error budgets now stop runaway retry loops
- Remote embedding verification now works with LiteLLM providers that return dictionary-style embedding rows (including Ollama-backed setups)
- Output Language now applies reliably across Feed memories, auto-capture, graph summaries, extension distill, and entity extraction
- Custom Instructions now apply to browser extension auto-capture and thread distillation, not only Feed/AI Now
- Linux terminal access controls are more consistent: TUI/CLI now manage same-network bind/allowlist and startup respects saved LAN settings
- Manual thread source filtering now behaves consistently across lists, detail, export, search, MCP tools, and background retrieval
- Large remote/browser Library uploads now stream more safely with less memory pressure, improving reliability for bigger documents
- Remote folder import flow now matches desktop behavior more closely, including summary events and duplicate detection
- Temporal extraction and distillation saves are more reliable: date-bound values and entity temporal fields now persist correctly
- One-click plugin install/update from Integrations now resolves PATH more reliably in desktop app environments
- AI Now now reads your own history more reliably: broad prompts pull from recent threads plus Library sources, not just Working Memory
- AI Now retrieval UI is cleaner and more coherent: thread/library results render as dedicated cards with calmer visual consistency
- AI Now mode routing is more stable with fewer scroll jumps, less redraw noise, and fewer unnecessary pre-search clarification turns
- AI Now Obsidian vault resolution is more robust: direct vault-name candidates are tried before slower full-vault listing
- AI Now no longer repeatedly re-opens Obsidian on macOS while it is already running
- Generic MCP clients now receive a smaller exported tool schema on the wire, reducing parser failures with strict grammar-based stacks
- Timeline typing preview no longer triggers excessive background search churn or restart-like behavior under rapid input
- Distillation now works on Linux and Windows without Pro: any configured remote LLM can handle distillation when local MLX is unavailable, and the CLI defaults to remote extraction on these platforms
0.6.132026年3月24日
Auto-Learn and a Better Context Picker
0.6.13 makes Mem quicker to use day to day. Library items can learn automatically after indexing, you can pull existing memories, sources, or threads into a prompt more directly, and Claude Desktop reconnects with less friction.
亮点
- Turn on Auto-learn in Settings → Memory Processing, and new sources extract knowledge automatically after indexing — no more clicking Learn one by one
- Attach sources, memories, or past conversations when composing in the Feed — search across all your knowledge, pick what's relevant, and ask questions grounded in context
- AI Now now reads your own history more naturally: broad prompts can pull from recent threads and show cleaner thread/library cards instead of raw tool JSON
- AI Now sessions feel calmer under heavy tool output, with less redraw noise and fewer unnecessary follow-up questions before first lookup
- Claude Desktop extension now installs on macOS and Windows with no separate Python or Node setup
- OpenClaw long-running sessions now sync reliably — large repetitive thread backups no longer fail on oversized CLI payloads
- Mobile and web apps detect when your server goes offline and show a clear status screen instead of a technical error message
- On macOS, you can now keep Mem in the menu bar after closing the main window; the Dock icon returns automatically when you open the app again
- Labels now propagate to crystals, merged knowledge, and search results — label-based filtering works reliably everywhere
- Browser extension capture is now much smarter: Manual Distill can create or update multiple memories from one conversation, and browser-created memories now keep a reliable link back to the full thread
新增
- Auto-learn sources: an opt-in toggle in Settings → Memory Processing that automatically extracts knowledge from new sources after indexing. Manual Learn always takes priority over auto-learn tasks
- Context picker in the Feed: a new button in the Feed toolbar lets you search and attach Library sources, memories, or past conversations when composing. Selected items appear as chips and the resulting memory links back to them
- Labels on ingestion: the upload API, file-path ingestion, and batch endpoints now accept labels — tag sources at the point of capture from the CLI, integrations, or the API
- Claude Desktop extension now installs on macOS and Windows with no separate Python or Node setup. For remote access, it reads the same shared config file as the `nmem` CLI
- New macOS Menu Bar & Dock preference: after you close the main window, Mem can stay in the top menu bar instead of remaining in the Dock. When you reopen the window, the Dock icon comes back automatically
- The `nmem` CLI is now automatically configured when the desktop app generates an API key — no more manually copying credentials
- Connect your phone in seconds: the desktop app can now show a QR code in Access Anywhere settings. Scan it from the mobile or web login screen to connect instantly — no more typing long URLs and API keys
- `nmem status` and the TUI now show whether you're connected locally or to a remote server — no more guessing which Mem you're talking to
- Sources in the Library can now be organized with labels — the same label system used for memories now extends to sources, so you can tag, filter, and group your reference material
修复
- AI Now now uses recent conversation threads and correct Library source-reading tools more reliably, so broad history questions don’t depend only on Working Memory
- Thread and Library retrieval results now render as calmer dedicated cards instead of raw JSON-like output
- Memory, graph, community, crystal, Working Memory, and thread cards now use more consistent spacing and headers across AI Now
- Memory mode and plan/graph card routing are more stable, with fewer scroll-position jumps during graph-heavy conversations
- Reduced accidental hot-path logging and unnecessary graph redraws make long tool-heavy AI Now sessions less flickery
- Obsidian vault lookup is less likely to stall: direct vault-name candidates are tried before full vault listing
- Browser extension Manual Distill now reads conversations more intelligently instead of flattening them into one quick save — it can inspect the thread selectively, check what already exists in Mem, and create or update multiple memories when that leads to a better result
- Browser-created memories now keep stronger provenance — when the extension writes a new memory, it first makes sure the conversation is backed up as a canonical thread in Mem, so you can trace where that knowledge came from later
- Labels now carry through to crystals and merged knowledge — when Background Intelligence distills or merges memories, the labels from source memories come along. Label-based filtering is now reliable across all knowledge types
- Memory search results now include labels — integrations like Claude Code and Cursor can see and filter by labels in search results
- Mobile and web clients no longer show a raw "Load failed" error when the server is unreachable — a clear reconnection screen appears instead, with automatic retry every few seconds
- If the server goes down while you're using the app, you're taken to the reconnection screen rather than seeing broken views — the app resumes right where you left off once the server is back
- Desktop update checks are now platform-aware. If a new build is available for macOS first, Windows and Linux devices keep seeing their own newest available build instead of a version they cannot install yet
- Thread source filters now work across your whole conversation history. The source menu lists every imported source, and selecting one refreshes the full paginated thread list instead of filtering only the current page
- The Threads source menu now scrolls cleanly when you have many sources, and icons like Gemini CLI, OpenClaw, Antigravity, Grok, Manus, and OpenWebUI render correctly
- Custom providers (SiliconFlow, together.ai, and other OpenAI-compatible services) now work reliably in AI Now — previously they could pass the connection test in Settings but fail when you started a conversation
- AI Now no longer hangs on startup when the agent process fails to launch — the error now appears immediately with diagnostic details instead of an infinite loading spinner
- GitHub Copilot with Codex models (like gpt-5.1-codex-mini) now works in AI Now — these models are automatically routed to the correct API endpoint
- GitHub Copilot headers updated across all integrations to match the current API, fixing authentication errors that appeared with newer Copilot releases
- ChatGPT Subscription models now work for background knowledge extraction — previously the request was silently misrouted and fell back to local processing
- OpenRouter models no longer fail with 400 errors during knowledge extraction — the provider is now correctly identified so requests go to the right API endpoint
- The desktop app no longer gets stuck on a white screen during startup when the loading view hands off to the main app
- The `nmem` CLI no longer shows a cryptic "Resource doesn't exist" error when the server is unreachable — it now explains the problem clearly and, for first-time remote users, shows exactly how to set up `~/.nowledge-mem/config.json`
- Server URLs with a trailing slash (e.g. `https://mem.example.com/`) no longer cause silent connection failures across the CLI, TUI, and all integrations
- OpenClaw long and repetitive sessions now back up reliably — thread sync no longer tries to squeeze whole conversation histories through one oversized CLI command, so repeated cron-style captures keep syncing instead of failing mid-session
0.6.122026年3月17日
Crash-Loop Recovery, VPN Connectivity & Mobile Fixes
Fixes crash-loop recovery, improves remote connectivity for Tailscale and VPN users, and resolves Android mobile app installation issues.
亮点
- The app now recovers automatically after a hard crash instead of getting stuck in a restart loop
- When your database needs more memory, the app detects and adjusts the allocation on its own — a brief notice on the startup screen lets you know
- macOS desktop app now connects to servers over Tailscale and other VPN networks without connection errors
新增
- Export and import now work from any device — back up or restore your knowledge base from a remote desktop, web browser, or mobile app without needing access to the server filesystem
- OpenClaw plugin: captured conversation messages can now keep more context — a new Max thread message chars setting (default 800) lets you raise the limit for code reviews and technical discussions that need fuller message capture
- OpenClaw plugin now reads remote credentials from the shared config file (~/.nowledge-mem/config.json) — set up Access Anywhere once and the OpenClaw, Bub, Claude Code, and nmem CLI all connect automatically
修复
- The app no longer gets stuck in a restart loop after an unexpected shutdown — leftover lock files are now cleared automatically on next launch
- Large knowledge bases no longer cause a startup crash loop when the database outgrows its default memory allocation — the app detects the problem, recovers safely, and automatically increases the allocation for next launch. The startup screen shows the adjustment and guides you if further tuning is needed
- macOS desktop app can now connect to servers on Tailscale and other VPN networks — Apple's security policy blocked HTTP to non-standard IP ranges like 100.x.x.x, which is now allowed
- Android mobile app can now connect to servers over HTTP — the release build was silently blocking non-HTTPS connections, preventing LAN and VPN connections
- Background agent no longer garbles URLs and CJK text in memories — search index tokenization was leaking into agent-visible content
- Bub plugin: AI responses are now faster when memory recall is active — context loading no longer forces the model to reprocess the full conversation from scratch on every turn
0.6.112026年3月16日
iPhone Remote Access and Smarter Remote AI
0.6.11 brings iPhone access to a remote Mem server, lets AI Now use the provider already configured on that server, and fixes a wide set of reliability issues across CLI auth, MCP transport, and provider behavior.
亮点
- Nowledge Mem now runs on iOS and Android as a native mobile app: connect to your always-on Mem instance and use your knowledge base from anywhere
- AI Now works across devices in remote mode: it uses the remote server's provider config automatically, with clear "Local AI" indicator in the UI
- New `nmem key` command lets you view or rotate Access Anywhere API keys directly from terminal
- Remote workflows are safer: remote file uploads, remote license actions, and Linux headless service-user state are now consistent
- Droid integration is now available as a first-class plugin path
新增
- iOS and Android mobile apps: connect to a desktop/server Mem instance through Access Anywhere with a native mobile experience. iOS via TestFlight, Android via APK download
- New `nmem key` CLI command for Access Anywhere API key retrieval and rotation (`nmem key`, `nmem key --rotate`)
- Remote-mode AI Now now uses the provider configured on the connected server automatically, so you don't need to set keys on every device
- Droid plugin integration added for Working Memory, routed recall, distillation, status checks, and resumable handoff summaries
- Provider context-window configuration is now available in Settings, CLI, and TUI for better small/large-model compatibility
- Custom OpenAI-compatible providers now support both Chat Completions and the newer Responses API
改进
- Access Anywhere controls moved into a dedicated Settings tab for clearer tunnel, key, and binding management
- Update bubble now opens a compact review panel with version info, changelog link, and install action
- `nmem status` now shows the CLI version separately from the server version and warns in WSL when localhost is probably reaching the Windows app instead of the Linux service
- OpenClaw plugin health checks and error reporting are now more reliable
修复
- Windows AI Now startup now recovers from corrupted session-state files instead of failing at launch
- Agent-created memories now preserve non-Latin text correctly instead of storing escaped Unicode sequences
- Distillation preview-to-save flow no longer fails after long review delays; cache lifetime is extended with fallback regeneration
- Large document indexing is more stable: long PDFs no longer stall ingestion for hours
- Remote client file ingestion now sends actual file content correctly instead of creating empty sources
- Headless Linux service installs no longer split state between user and `/root`; existing installs should run `nmem service install` once to realign
- Remote license actions (activate/deactivate/renew) now target the connected backend consistently
- If an older Linux install left Mem state owned by `root`, license actions, provider saves, and Access Anywhere settings now explain which path needs its ownership fixed instead of failing with a generic API error or silently reverting later
- On Linux, stateful `nmem` commands now stop if you run them as `root`, so a fresh install is less likely to leave `root`-owned config behind by accident. Use `nmem --allow-root ...` only if you intentionally want a root-owned deployment
- Background tasks now always emit visible timeline results, including no-op and completed states
- Library folder import limits and edge cases are fixed: cap raised to 10,000 files and folders ending with `.md` are no longer misclassified as files
- "Fetch models" now works correctly when using a saved API key with a custom base URL
- AI Now model selection no longer resets when you edit provider settings
- OpenRouter models under the `openrouter` organization (e.g., `openrouter/auto`) no longer fail with invalid model ID errors
- Custom Responses API providers no longer fail verification due to input format mismatch
- Context window values entered in Settings now persist correctly across page reloads
- Default provider selection now persists when switching between Settings tabs
0.6.102026年3月13日
Windows In-App Updates and Better Remote Access
0.6.10 brings Windows in-app updates to parity with macOS and tightens remote access across web and CLI flows. It also fixes the rough edges around mobile-browser remote access.
亮点
- Windows updates are now in-app: the installer downloads in the background and a single click installs and relaunches — no more opening a browser to download manually
- Windows uninstaller now shuts down the app cleanly before removing files, instead of showing a force-kill dialog
- Access Anywhere web entrypoint at `/app` is more reliable with a clear fallback when unavailable
改进
修复
- Opening `/app` no longer returns a raw 404 when web assets are missing
- Web sidebar toggle no longer has a macOS-only offset in browser mode
- macOS CLI uninstall now correctly reports when the admin prompt is cancelled
- Windows installer no longer fails to update PATH when the install directory contains spaces
- Claude Code plugin hooks in WSL now auto-bridge to Windows `nmem.cmd` when native `nmem` isn't available
0.6.92026年3月12日
Remote Mem from Desktop, Web, and Raycast
0.6.9 makes remote access feel like a real product surface. The desktop app can connect to a remote Mem as a client, Raycast can reach your Mem from anywhere, and the Graph API plus Windows startup behavior both get sturdier.
亮点
- Desktop app can now run as a client to connect to a remote Mem app/server, so you can use the full desktop UI against a remote knowledge base
- Access Anywhere now includes a web client — open your Mem URL with /app from any browser to search, browse, and explore your knowledge graph
- Raycast extension now supports Access Anywhere too, so you can search and explore a remote Mem app directly from Raycast
- New embeddable Graph Visualization API enables external apps to explore your knowledge graph
- Windows startup flow is safer: installer PATH handling and AI Now session management are hardened
- WSL support: coding agents like Claude Code running in WSL can now use nmem directly via Windows interop
- Bub plugin: cross-tool knowledge flows between Bub and all your other AI tools
新增
- Desktop app client mode is now supported for remote Mem: connect from one desktop app to another Mem server/app and work with remote memories using the native desktop interface
- Access Anywhere now supports browser extension and web browser workflows — open your Mem URL with /app to search, browse memories, and explore the graph from any device
- Raycast extension now supports Access Anywhere for remote Mem, so you can query and navigate remote knowledge without opening the desktop app window
- New Graph Visualization API for embedding Nowledge Mem graph exploration into other applications
- Shared HTTP graph exploration endpoint `GET /graph/explore` provides visualization-ready neighborhoods from one or more memory IDs
- Bub plugin: knowledge from Claude Code, Cursor, ChatGPT, and other tools is now searchable inside Bub — and what you learn in Bub flows back to every other tool
改进
- Access Anywhere settings are clearer: setup is split into "Connect to Another Device" (client mode) and "Share This Device" (server mode), with consistent naming across the UI
- Raycast graph exploration now uses the same backend `/graph/explore` contract as documented HTTP clients, improving consistency across clients
修复
- macOS: `nmem` CLI now installs to `/usr/local/bin` so it works immediately — no PATH setup or new terminal needed
- Windows installer no longer risks corrupting user PATH by expanding `%VARIABLE%` entries into literal values during install
- AI Now session resume/cleanup on Windows no longer depends on Git Bash for session-state helpers
0.6.82026年3月10日
Lower Memory Footprint, Smaller Indexes & Startup Hardening
This release dramatically cuts memory and disk pressure for long-running desktop usage, while fixing reliability issues in deep-learn, background task retry, Access Anywhere readiness, and Windows startup flows.
亮点
- Idle memory drops significantly with automatic model unload and tighter cache caps, reducing pressure on 8 GB machines
- Search index size drops sharply with faster old-version pruning and on-demand Optimize controls
- Reindex now uses much less peak RAM, making large-library rebuilds safer on typical laptops
- Library deep-learn now works more consistently across providers, including stronger fallback behavior
新增
改进
- Idle memory footprint is much lower: search model memory is reclaimed after inactivity and backend caches are capped
- Search index storage is smaller by default: old data versions are pruned faster, with manual optimize available when needed
- Reindexing large libraries now runs with dramatically lower peak memory and cleaner post-reindex recovery
- Raycast graph exploration now shares the same backend `/graph/explore` contract as documented HTTP clients
修复
- Library deep-learn now works reliably across providers, with structured extraction paths instead of provider-dependent behavior
- Feed "save this as memory" requests now persist correctly instead of returning a summary without writing data
- Retry button now appears consistently for failed/aborted background tasks
- Access Anywhere on Linux/headless now waits for true tunnel readiness and retries transport when needed, reducing false-live states
- Fresh server installs now use the correct modern database naming (`nowledge_graph_v2.db`)
- AI Now Windows startup is more resilient to PowerShell/source-encoding edge cases
- Windows PATH corruption risk is removed from installer flows by preserving expandable registry semantics
- AI Now session management on Windows no longer depends on Git Bash being installed
0.6.72026年3月10日
Sharper Integrations and Safer Access Anywhere
0.6.7 makes setup less brittle across Gemini CLI, Cursor, browser capture, and remote thread save paths, while improving everyday diagnostics and import reliability.
亮点
- Gemini CLI is now live in the official Extensions Gallery, and both Gemini CLI and Cursor now have clearer setup guides and polished integration landing pages
- Remote-safe thread save now works through local session parsing for supported coding agents instead of relying on server-side agent-path access
- Secure multi-device access now uses API keys by default, with Copy and Rotate controls
- Import entire folders into Library or Feed in one step, with clear added, skipped, and error summaries
- AI failures now include inline diagnostics with Details and Copy diagnostics for faster troubleshooting
新增
- New Gemini CLI integration path with hooks, Working Memory, routed recall, real session import, distillation, resumable handoff support, and a live official gallery listing
- New Cursor plugin path with bundled rules, skills, MCP wiring, and a smoother marketplace-first setup flow
- New website integration surfaces for MCP, Chrome Extension, npx Skills, and CLI so first-time users can choose a setup path faster
- OpenAI Responses API support for custom providers, selectable per endpoint in Settings
- Folder import for Library and Feed: recursively scans supported documents while skipping common build and system directories
- DeepSeek conversation import from chat.deepseek.com export data
- Universal conversation markdown import format with optional metadata frontmatter
- New `nmem t import` CLI command: import from markdown, JSON, or stdin with format auto-detection
- Secure LAN access defaults: generated API key, Copy LAN Setup block, and one-click key rotation
- Quick log access from startup failures and Settings > Info with one-click open-log-folder actions
- `memory_search` now returns `confidence` (0.5-1.0) per result for MCP clients
改进
- Integration pages now use clearer product language, calmer visual hierarchy, dedicated command blocks, and more useful first-time setup guidance
- Search ranking now accounts for memory confidence from real usage signals like search, open, and read, helping well-validated memories surface earlier
- Automatic labeling now applies on creation and reuses your existing label vocabulary for more consistent organization
- AI Now `@` picker now spans memories, threads, and Library sources in one place
- Large file workflows in AI Now are faster: the agent now prefers search and sample-first analysis before full reads
- Library status language is clearer: "Searchable" versus "Deep-learned" with filter support and batch deep-learn actions
修复
- Thread save reliability improved: `nmem t save --from claude-code|codex|gemini-cli` now parses local session files without requiring server-side agent-path access
- Gemini CLI now separates full thread save from resumable handoff summaries, making capture behavior clearer during long coding sessions
- Subscription status and premium gating now use one unified access contract across Plans, AI Now, Remote LLM, and backend checks, reducing misleading states
- Windows second-launch boot failure fixed for 0.6.5 and 0.6.6 upgrade paths, including stale CLI wrapper and stale-port recovery handling
- AI Now now surfaces provider and startup failures inline and keeps chat context visible, with copyable diagnostics instead of silent hangs
- Search results now show correct confidence values, separate from relevance score, and agent-assigned labels are visible consistently across surfaces
- Plugin toggles in AI Now now actually filter tools, including custom MCP tool dedup fixes
- ChatGPT Subscription and GitHub Copilot OAuth flows now work correctly in the CLI and TUI test-connection flow
- Memory distillation no longer fails to blank output silently: it falls back to the local LLM or returns clear failure details
- Browser extension reliability updates: Grok detection on x.com and improved unsupported-browser messaging for failed AI-access checks
平台
- Linux package uninstall parity: both `nmem` and `browse-now` symlinks are now cleaned up consistently
- Startup path resolution is more resilient across macOS, Windows, and Linux after updates, including plugin, CLI, and Python fallback paths
0.6.62026年3月4日
Windows Startup Hotfix and Provider Reliability
0.6.6 removes a Windows startup blocker, improves Intel Mac compatibility, and makes AI Now plus remote-model workflows less fragile when providers misbehave.
亮点
- New experimental Data Transfer: export a portable snapshot and import it later with merge/skip/overwrite modes and automatic backup
- Critical Windows PATH corruption fix: startup no longer rewrites expandable PATH entries and breaks other CLI tools
- Windows idle GPU burn is eliminated by making the aurora background static
- Startup reliability is stronger across Linux/macOS/Windows, including RPM runtime self-healing and safer DB path recovery
- Obsidian integration now uses official CLI workflow with robust fallbacks for real-world vault edge cases
- Claude Code plugin v0.7.0: full lifecycle hooks for Working Memory, session capture, and per-turn search/save nudge
新增
- Data Transfer (Experimental): export a portable snapshot and import into another install with merge/skip/overwrite options, selective inclusion toggles, and a safety backup before import
- Claude Code plugin v0.7.0: Working Memory loaded at session start/resume/clear, async session capture after every response, and per-turn behavioral nudge with search/save syntax
- nmem-cli config file support: set remote connection once in ~/.nowledge-mem/config.json (flags > env > config > defaults). `nmem status` now shows where values came from.
改进
- Windows GPU usage is now near-zero at idle: aurora background is static by default, removing continuous recomposition load
- AI Now and secondary surfaces now use a more consistent Phosphor icon system with unified sizing/weight behavior
- v0.6 UI/ICON polish across thread detail cards, glowing cards, and pagination for cleaner visual hierarchy
- Obsidian setup guide now uses full-screen screenshot preview for faster first-time setup
- Obsidian plugin now uses official CLI tools as source of truth for better multi-vault consistency
修复
- Windows startup no longer corrupts user PATH variable types (`REG_EXPAND_SZ` preserved), preventing breakage of tools like pnpm/node/git
- Removed false-positive VC++ startup guidance so unrelated failures no longer send users into unnecessary reinstall loops
- Provider model auto-fetch loop in Settings is fixed: empty/unsupported model endpoints no longer trigger repeated retries
- LLM provider adaptation is more robust for non-standard model-list APIs: one-time auto-attempts are deduplicated, while manual model entry remains available as a stable fallback
- Apple Notes search now shows actionable error states instead of misleading "No notes found" when tool/runtime calls fail
- Obsidian reliability fixes: resilient note parsing, text-search fallback when CLI returns zero hits, and safer vault discovery under transient CLI failures
- Tool renderer syntax/JSX structural issues are fixed, resolving rendering instability in slide and notes integrations
平台
- Linux RPM upgrades no longer remove bundled runtime during upgrade; startup can self-heal from packaged runtime archive if needed
- Config/data paths are standardized: app config/state in AppConfig (`co.nowledge.mem.desktop`) and graph DB/log data in AppData (`NowledgeGraph`) with compatibility migration
- Startup DB path recovery now prefers existing `nowledge_graph_v2.db` when stale overrides point to empty/unversioned paths
0.6.52026年2月28日
Windows Startup Recovery & Intel Mac Reliability
This hotfix removes a startup blocker on some Windows machines, improves Intel macOS compatibility, and tightens provider reliability so AI Now and remote-model workflows fail less often and recover faster.
亮点
- Known issue (0.6.5): AI Now startup can fail on macOS, Windows, and Linux; dedicated hotfix is in progress
- Windows users now get a direct in-app fix path when Microsoft VC++ runtime is missing
- Intel macOS 15+ startup checks and search embedding path now work consistently
- MiniMax Anthropic endpoint defaults are safer, reducing AI Now 404/provider mismatch failures
- Configured providers can now be removed safely in both Desktop Settings and `nmem` CLI/TUI
修复
- Windows startup no longer dead-ends on missing VC++ runtime (`_lbug` DLL load): Startup Flow now shows clear remediation links and the same guidance is mirrored in logs for support teams
- Intel Macs on macOS 15+ are no longer incorrectly blocked before backend initialization
- Intel macOS search embeddings now follow the correct non-Apple-Silicon path end-to-end, improving model download, verification, and reindex stability
- Local LLM capability detection on macOS is now architecture-aware: Intel Macs correctly guide users to Remote LLM workflows
- MiniMax provider defaults now favor Anthropic-compatible endpoints and remove incompatible presets that caused cross-surface failures
- Moonshot CN K2.5 extension tool-loop requests now use valid payload semantics, fixing 400 errors in capture workflows
- Token budget settings via `nmem config settings set` now take effect immediately — previously the API silently dropped these values so only GUI changes worked
改进
- Provider lifecycle is cleaner across GUI and terminal workflows: delete providers safely and let purpose fallback resolve automatically
- Intel macOS search model metadata is now shown with correct platform/backend expectations
平台
- Known issue: AI Now startup regression can prevent AI Now session boot on all platforms in 0.6.5; dedicated hotfix is in progress
- Bundled Ladybug fork is rebased to upstream 0.15.0 while preserving desktop-specific integration behavior
0.6.42026年2月28日
Smarter Tables and Faster Long Lists
0.6.4 turns CSV and XLSX files into something you can actually work with: charts, cleaner table rendering, and better searchable chunks. It also makes large libraries and long timelines easier to browse, with more provider control from the CLI.
亮点
- CSV/XLSX now auto-generate interactive charts in Feed and Library
- Timeline now loads in larger batches and can jump to any calendar date on demand
- Library source list now paginates progressively, keeping large collections responsive
- Agent provider/model purpose can now be configured fully from `nmem` CLI/TUI
- DeepSeek is now supported across desktop settings, AI Now, CLI/TUI, and extension flows
- Linux `.deb` installs now support signed APT updates and safer upgrade shutdown
- Apple Silicon idle memory drops significantly with automatic model unload when inactive
- OpenAI-compatible local Embeddings API for tools and scripts
新增
- OpenAI-compatible local Embeddings API (`/v1/embeddings`, `/v1/models`) for external tools and automation
- Agent source management: Feed and Background Intelligence can now discover and delete Library sources safely in one workflow
- New `nmem sources` CLI commands (`list`, `show`, `delete`) for source lifecycle control from terminal workflows
- Terminal-only provider purpose control: `nmem` CLI/TUI can now set default provider and Agents (`ai_now`) purpose provider/model without opening GUI Settings
- DeepSeek provider support is now end-to-end across app settings, `nmem` CLI/TUI, AI Now, and browser extension flows
- Tabular intelligence: auto chart detection plus schema-based chunking makes CSV/XLSX content more searchable and answerable
- Claude Desktop extension (v1.3.0): configure Remote URL + API Key directly in extension settings for Access Anywhere
- Library pagination: source list loads progressively with "Load more", keeping large libraries fast
- Timeline "Load more": feed loads 200 items initially and grows on demand, with calendar-aware date fetching for any historical day
- Linux signed APT repository for `.deb` users: one-time setup enables normal `apt upgrade` updates
- Safer Linux package upgrades: app now checkpoints and shuts down gracefully before package replacement to protect database state
- AppImage auto-update now supports download, install, and restart directly in-app (same flow as macOS)
改进
- Search remains responsive during heavy Background Intelligence write bursts with per-table index rebuilds and debounce scheduling
- Table reading experience upgraded across Feed, Library, and AI Now with sticky headers, better scrolling, and clearer numeric alignment
- Feed Agent table analysis now reads and reasons in a single pass, producing clearer analytical answers and visualizations
- Library Chat keeps QA history when switching files and across window reloads, with per-source isolation
- Timeline calendar click now navigates to any date, fetching just that day's events on demand instead of requiring all items pre-loaded
- Windows WebView graphics usage reduced by removing redundant animated layers and pausing animations when app is not visible
- Apple Silicon memory behavior improved for long-running sessions: inactive local models auto-unload and reload when needed, reducing idle RAM
修复
- Feed Agent context amnesia fixed: recent sessions persist (time/turn bounded) and timeline summaries are injected for continuity
- Source deletion dead-ends resolved: agents can now find source IDs, handle parse-failed sources, and avoid follow-up loop explosions
- Token budget false positives removed: budget checks now match real context usage and limits are configurable in Settings and CLI
- Large file preview no longer freezes Library UI; previews are capped and rendered progressively
- macOS updater no longer hard-kills the backend process, preventing Kuzu WAL corruption risk during install
- macOS update confirmation now uses native Tauri dialog for reliable user consent before install
- TUI provider selector no longer crashes on configured custom provider IDs due to Select markup parsing
- Moonshot CN K2.5 tool-loop reliability improved in browser extension: force-capture now uses Moonshot-native thinking disable semantics instead of invalid reasoning-effort values
- Vega-Lite charts no longer render at zero height in Feed and AI Now responses
平台
- Embeddings model transparency: `/v1/models` now reports one canonical backend model per platform to avoid alias ambiguity
- Embeddings requests now require canonical model IDs, preventing silent model mismatch in automation clients
- macOS local LLM download is now optional; remote LLM-first workflows are fully supported
0.6.32026年2月26日
Apple Notes, More Capture Sources, and Better Windows Startup
0.6.3 brings Apple Notes into AI Now, lets you connect multiple Obsidian vaults and Notion workspaces, adds display zoom, and makes Windows runtime setup more dependable.
亮点
- Apple Notes: search and read your notes directly from AI Now
- Multi-instance: connect as many Obsidian vaults and Notion workspaces as you need
- Display zoom: scale the entire interface to your preference
- OpenClaw plugin: trace any memory back to the full conversation it came from
- Alma plugin: status diagnostics and live settings — change credentials without restarting
新增
- Apple Notes plugin for AI Now: search and read notes from macOS Notes.app — no export needed
- Multi-instance Obsidian and Notion: connect multiple vaults or workspaces, each with independent config and auth
- Display zoom in Preferences: Small, Default, Large, or Extra Large — scales the whole interface
- Token budget controls in CLI and TUI: view and set hourly/daily limits from the terminal
- OpenClaw plugin 0.6.5: thread search and progressive message retrieval across 9 tools — trace memories back to full conversations
- Thread API pagination: fetch specific pages of long conversations instead of loading everything at once
- Alma plugin 0.6.3: status diagnostics, live settings reload, and Access Anywhere support — 12 tools total
改进
- Notion plugin handles unexpected parameters gracefully and validates tool schemas more robustly
- Plugin context uses uniform block tags — cleaner message rendering across all plugins
- Agents model routing: OpenRouter tool-call validation and model ID normalization
- npx Skills and Codex prompts updated with structured memory types and labels
修复
- Windows Store: bundled Python runtime works reliably under MSIX packaging
- Downloaded models found correctly on all platforms after cache path alignment
- TUI: embedding model status no longer shows "Error - None"
- Browser extension: Gemini folded messages and collapsed content captured correctly
平台
- Windows: CLI wrappers (nmem, browse-now) installed to user PATH for Store and non-Store builds
- macOS: Apple Notes integration — read-only, iCloud-aware
0.6.22026年2月25日
Doubao, WSL2 & Smarter Capture
Sync Doubao conversations, discover WSL2 sessions on Windows, and enjoy dramatically lower background costs with intelligent capture controls.
亮点
- Doubao (豆包) conversations now sync through Exchange
- Session Browser finds conversations inside WSL2 on Windows
- OpenClaw autoCapture uses up to 95% fewer tokens when idle
- Imported threads show clean tool summaries instead of raw output
新增
- Doubao (豆包) support in Exchange: sync and export your conversations from doubao.com
- WSL2 session discovery on Windows: Session Browser and Auto-Sync now find Claude Code, Codex, and OpenCode sessions inside WSL2 distributions
- Multi-instance Obsidian & Notion plugins: connect multiple Obsidian vaults or Notion workspaces simultaneously, each with independent config and auth
改进
- Imported threads show concise tool summaries instead of raw grep results and file dumps
- MCP Live Traffic in Integrations now displays real calls and responses
修复
- OpenClaw plugin: autoCapture no longer burns tokens from idle heartbeat cycles (5-layer cost defense, ~95% reduction)
- OpenCode sessions re-import correctly after deletion from Session Browser
- MCP Live Traffic log parsing updated for current backend format
0.6.12026年2月24日
Cost Controls & 20+ Languages
See what background intelligence costs, set your own limits, and let the system pause before you overspend. Knowledge augmentation now works in 20+ languages.
亮点
- Token budget with hourly and daily limits, plus a live usage dashboard
- Knowledge augmentation in 20+ languages including Chinese, Japanese, Korean, French, and German
- Thread viewer handles 4,000+ messages without lag
- macOS reindex uses up to 97% less GPU memory
新增
- Token budget controls: set hourly and daily spending limits for background intelligence
- Live token usage in Settings, the status API, and nmem status CLI
- Knowledge augmentation in 20+ languages: insights, briefings, and extractions in the language you choose
- Kimi Coding Plan provider support
改进
- Thread detail view handles 4,000+ messages smoothly
- Reindex on Apple Silicon uses up to 97% less GPU memory
- Smarter PDF detection: arXiv-style links and /pdf/ routes go through the file pipeline automatically
- Per-task safety net prevents single background tasks from running away
- Lower memory footprint during indexing on Windows and Linux
修复
- Preferred output language now persists correctly across restarts
- Browser Extension: Gemini table data captured correctly
- Feed: rapid consecutive submissions no longer dropped
- Feed: failed URL captures no longer create broken memories
- Settings: correct local model name shown instead of unrelated models
- Background Intelligence sessions recover gracefully after unexpected exits
- Direct PDF URLs from timelines now ingest through the file pipeline
0.6.02026年2月18日
Background Intelligence, Smart Feed & Library
Your knowledge now lives in a neutral layer that works with any AI tool. Switch freely between Claude Code, Cursor, ChatGPT, or whatever ships next. Background Intelligence builds connections and writes your daily briefing. Smart Feed, Library, and Exchange v2 round out the release.
阅读发布博客亮点
- Smart Feed: one input for captures, questions, and file imports
- Background Intelligence: the system finds connections, surfaces contradictions, and writes daily briefings on its own
- Working Memory: a daily briefing of what matters, delivered to ~/ai-now/memory.md
- Library: search PDFs, Word docs, and presentations alongside your memories
- Access Mem Anywhere: one-click secure remote URL + API key for your other machines and agents
- Exchange v2 (Chrome): auto-captures AI conversations across 13+ platforms
- Browse-Now (Chrome): give AI agents access to your browser
- Search engine rewritten in Rust with new BM25, vector search, and SOTA local embeddings
- Alma plugin: native memory integration with tools, command palette actions, and auto-capture controls
新增
- Smart Feed: one input for captures, questions, and file imports
- Unified timeline: captures, insights, flags, and daily briefings in one stream
- Background Intelligence: the system finds connections, surfaces contradictions, and writes daily briefings
- Working Memory: daily briefing at ~/ai-now/memory.md, readable by any MCP agent
- Memory evolution: detects when memories update, enrich, confirm, or contradict each other
- Library: search PDFs, DOCX, PPTX, and Markdown alongside memories
- Access Mem Anywhere: start a secure Cloudflare tunnel from Settings (Quick link or Cloudflare account mode)
- Remote API auth hardening: every remote request requires your API key, with key rotation support
- Exchange v2: auto-captures AI conversations across 13+ platforms (Claude, ChatGPT, Gemini, and more)
- Browse-Now CLI: give AI agents access to your browser with your login sessions
- Interactive graph visualization in Claude Desktop and other MCP hosts
- Chinese full-text search
- AI-powered search reranking with human-readable explanations
- Session auto-sync: real-time watching for Claude Code, Cursor, Codex, and OpenCode
- Linux headless server deployment with systemd
- Full CLI for server setup, license, LLM config, and knowledge settings
- New providers: ChatGPT (Codex) subscription, MiniMax, Z.AI, MoonShot AI
- MCP: Graph View in Memory Search for Claude Desktop and other MCP-UI hosts
- OpenClaw plugin: native memory integration for the OpenClaw agent framework
- Alma plugin: native memory integration with tools, command palette actions, and optional auto-capture
- Raycast extension: search memories, save insights, read Working Memory
- npx Skills: one-command memory integration for any agent (search, save, Working Memory, graph exploration)
- Knowledge structure: Trace → Unit → Crystal three-layer model inspired by cognitive science
- Community detection: automatic topic clustering across the knowledge graph
- Node importance scoring: surfaces the most influential concepts in your graph
- LLM-friendly docs: every page serves clean Markdown via Accept: text/markdown header, plus /llms-full.txt and /llms.mdx endpoints
改进
- Search engine rewritten: new tokenization, index engine, and SOTA local embedding models
- MCP: 24 tools exposing the full knowledge surface (up from basic search/add/labels)
- Visual refresh: calmer, cleaner design with content-first layouts
- Search is 50% faster with native Chinese support
- Graph timeline slider: filter your knowledge graph by time range
- Real-time streaming: see agent thinking as it happens
修复
- Graph visualization performance improved
- OAuth connections silently failing on macOS
- Search Index Model downloads failing on Windows
- Session export showing wrong message count from Cursor
- Access Mem Anywhere copy actions now survive desktop reload/restart without forcing key rotation
- Cloudflare account token parsing now accepts full command formats (`service install`, `--token`, `--token=`)
平台
- Linux: Full headless deployment with nmem serve and nmem service
- Linux: Interactive TUI control plane for headless server management
- Community: native support in DeepChat and LobeHub
0.5.112026年1月8日
MCP Browse Mode & CLI Thread Import
Browse recent threads and memories via MCP, and import Claude Code or Codex sessions directly from the CLI.
亮点
- MCP: Browse threads and memories without a search query
- CLI: Import Claude Code and Codex sessions with `nmem t save`
- Re-import conversations without duplicates
新增
- MCP list mode: browse recent threads and memories without a search query
- MCP source filter: filter threads by source (e.g., 'claude-code')
- CLI `nmem t save`:import Claude Code and Codex sessions as threads
- Memory search now shows which thread a memory came from
改进
- Claude Code skills now use nmem CLI for faster, more reliable operations
- Codex custom prompts updated to use nmem CLI
- Re-importing conversations appends new messages instead of failing
- Automatic deduplication when re-importing the same conversation
0.5.102026年1月4日
Standalone CLI & Batch Thread Deletion
Access your Nowledge Mem from any machine with the new standalone CLI package.
亮点
- Standalone CLI: pip install nmem-cli
- Remote access to your Nowledge Mem server
- Batch delete multiple threads at once
新增
- Standalone CLI package: access Nowledge Mem from any machine
- Batch thread deletion: select and delete multiple threads
- MCP thread deletion tool for AI integrations
0.5.92025年12月31日
AI Now Command Execution Fix
Fixes an issue with command execution in AI Now.
亮点
- Fixed AI Now command execution
修复
- AI Now command execution not working correctly
0.5.82025年12月30日
Internationalization & Infographic Visualization
Full English and Chinese language support. Infographic visualizations in AI Now.
亮点
- Full English & Chinese language support
- Infographic visualizations in AI Now responses
- Improved database reliability and performance
新增
- Internationalization: full English and Chinese translations
- Language switcher in Settings with automatic system detection
- Infographic visualizations for research and presentation skills
- Plugin auto-updates when AI Now starts
- macOS compatibility check for Apple Silicon and macOS 15+
改进
- Better detection and display of AI reasoning/thinking content
- Markdown formatting: fixed newlines and structured content
- Database reliability and data integrity on shutdown
- Faster and smoother UI updates
- Search and pagination performance
修复
- Pro user status showing incorrectly when offline
- Label changes being overwritten
- Various markdown rendering issues in AI responses
- Folder and file operation issues
- Data loss issues
0.5.72025年12月23日
AI Conversation Discovery
Find and import your conversations from Claude Code, Cursor, and Codex with a single click.
亮点
- One-click discovery of AI coding conversations
- Import from Claude Code, Cursor, and Codex
- Searchable model dropdown for 20+ models
新增
- AI Conversation Discovery: automatically find conversations from Claude Code, Cursor, and Codex
- Preview conversations before importing (project name, message count, last modified)
- Batch import: select and import multiple conversations at once
- Searchable model select for providers with many models
改进
- Model selection shows filtered/total count for easier navigation
- Faster parsing of large conversation files
0.5.62025年12月23日
16x Larger Memories & AI Now Stability
16x larger memory capacity, better AI Now stability with integrations, smoother UI.
亮点
- 16x larger memory capacity (up to 32,768 characters)
- Improved AI Now stability with integrations
- Smoother UI with stable pagination
新增
- 16x larger memories: up to 32,768 characters per memory
- GLM-4.7 and Gemini 2.5 Flash model support
改进
- AI Now stability with Notion and other integrations
- Better search for long memories
- Tool compatibility with Groq, OpenRouter, and more providers
- More reliable AI responses across providers
- Presentation and Travel Planner skill instructions
- Smoother pagination and loading states
修复
- Notion integration errors (400 Bad Request)
- Remote LLM toggle not taking effect
0.5.52025年12月22日
Arch Linux Support & Windows Fixes
Arch Linux support, improved link handling on Windows, and a What's New button in Settings.
亮点
- Arch Linux support, thanks to Hong Lan
- Fixed links not opening on Windows
- What's New button in Settings
新增
- Arch Linux package support
- What's New button in Settings → Info section
修复
- Links not opening on Windows
- Remote LLM configuration with custom provider only
平台
- Linux: Added AppImage standalone executable
0.5.42025年12月20日
Linux Packaging & Graph Reliability
Linux packaging improvements with AppImage and RPM support, plus knowledge graph reliability enhancements.
亮点
- AppImage support for most Linux distributions
- RPM packages for Fedora, RHEL, CentOS, openSUSE
- Orphaned entity cleanup in knowledge graph
- Bi-temporal relationship tracking
新增
- RPM packages for Red Hat-based distributions
- Orphaned entity cleanup tool
- Bi-temporal fields for relationships
改进
- Community detection algorithm reliability
- Entity alias matching and deduplication
- Faster installation
修复
- LLM provider handling for Gemini and Anthropic
- AI Now tool call display
- Linux AppImage archive paths
0.5.32025年12月18日
nmem CLI & New LLM Providers
New command-line interface for AI agents and power users, GitHub Copilot and DeepSeek R1 support, plus expanded conversation import options.
亮点
- nmem CLI for terminal-based memory access
- GitHub Copilot integration (no API key needed)
- DeepSeek R1 with visible reasoning traces
- ChatGPT & ChatWise ALL historical conversations import
新增
- nmem CLI with memory search, thread management, markdown import
- GitHub Copilot as LLM provider (uses existing subscription)
- DeepSeek R1 reasoning model with extended thinking
- Custom OpenAI-compatible endpoint support
- ChatGPT JSON export import
- ChatWise conversation import
- Bulk file import, import ALL historical conversations from ChatGPT & ChatWise(more upcoming)
- Message selection when distilling threads
- Source thread lookup for any memory
- Global keyboard shortcut customization
改进
- Obsidian tag extraction with Unicode support
- Wikilink parsing (header links, escaped pipes)
- Thread full-text search
- Automatic recovery from errors
平台
- macOS: Fixed parallel inference on Apple Silicon
- Windows: Fixed orphan process cleanup
- Linux: Fixed blank screen on Arch Linux
0.5.22025年12月9日
Deep Research Hotfix
修复
- Deep Research output parsing
- Research renderer confidence level handling
- Linux update download flow
0.5.12025年12月7日
Stability Hotfix
修复
- Parallel inference on Apple Silicon
- AI Now on Windows Store version
- Linux script path detection
0.5.02025年12月5日
AI Now & Bi-Temporal Memory
AI Now personal assistant, bi-temporal memory with cognitive-science decay model, Linux desktop support.
亮点
新增
- AI Now: Personal assistant with memory, files, and app integrations
- Notion, Obsidian, Apple Notes integrations for AI Now
- Deep Research with multi-agent web + memory search
- Document analysis (Excel, CSV, Word)
- Presentation generation with PowerPoint export
- Bi-temporal memory (event time vs record time)
- Memory decay model inspired by cognitive science
- Temporal intent detection for natural queries
- Auto-update system with progress tracking
改进
- Knowledge graph extraction up to 10x faster
- Fast search under 100ms
- Deep search up to 5x faster
- Graph visualization smoothness
平台
- Linux: DEB packages for Ubuntu/Debian
- Linux: RPM packages for Fedora/RHEL
- Linux: AUR support for Arch Linux
- Linux: Dedicated launcher UI
0.4.112025年11月18日
Light Mode
Complete light mode support with theme-aware visualizations.
亮点
- Light mode across all views
- Entity and community search
- Windows startup improvements
新增
- Light mode with theme toggle in Settings
- Entity search in fast search
- Community search across graph
- Label filtering in fast search
改进
- Theme-aware graph visualization
- LLM availability error reporting
- Settings UI with tooltips
修复
- Windows launcher startup flash
- Windows console window blink
0.4.102025年11月15日
Windows Beta
Windows support arrives with all core features working.
亮点
- Windows Beta release
- Light mode theming
- Optimized Deep Search (~6 seconds)
新增
- Windows Beta with full feature support
- Light mode with system theme integration
- Cross-platform embedding support
改进
- Deep Search optimized to ~6 seconds
- Intent detection accuracy
- Process cleanup on Windows
0.4.62025年10月29日
DeepChat & MCP Improvements
Native DeepChat integration and improved MCP tool definitions.
亮点
- Native DeepChat integration
- ChatHub conversation import
- Optimized MCP tool prompts
新增
- Native DeepChat integration
- ChatHub conversation import
- Entities API endpoints
改进
- MCP tool argument prompts
0.4.22025年10月24日
Thread Search & Claude Code
Search across all threads and import from Claude Code and Codex.
新增
- Thread search across all conversations
- Claude Code plugin for conversation import
- Codex integration with thread import
改进
- Thread view performance
- Code block rendering for large threads
0.3.12025年10月12日
Fast Search & Remote LLM
Quick memory lookup and cloud LLM provider support.
亮点
- Fast search mode for quick lookup
- Remote LLM provider support
新增
- Fast search mode
- Remote LLM provider support
改进
- LLM-based intent detection
0.2.42025年9月22日
License & Updates
Pro license activation and update checker.
新增
- Pro license activation
- Update checker
- Manual memory creation
- Discord and docs links in Settings
0.2.02025年8月7日
Alpha Release
First public alpha with core memory management, knowledge graph, and MCP integration.
亮点
新增
- Threads view for conversation management
- Memories view with knowledge cards
- Knowledge graph visualization
- MCP server for Claude, Cursor, and other AI tools
- Browser extension (Nowledge Mem Exchange)
- Search launcher (⌘⇧Space)
- Semantic search with embeddings
- Graph multi-selection and lasso select
- Native macOS app