Kynoptic records which apps you use and what your machine was doing into one local file that Claude, Cursor, or any MCP AI can search. So next time you can't remember where you saw or did something — just ask.Kynoptic 在后台记下你用过哪些软件、机器当时在干什么，存成一个 AI 能查的本地文件。下次想不起某个东西在哪看过、哪天做过，直接问就行。
# What was I doing on the computer yesterday between 2 and 4 pm?我昨天下午两点到四点在电脑上干了什么？ According to Kynoptic:根据 Kynoptic 的记录： 14:02–15:10 VS Code — repo "kynoptic", 3 files, deep-focus session14:02–15:10 VS Code —— kynoptic 仓库，3 个文件，深度专注 15:10–15:26 Chrome · 23 tabs — MDN, GitHub issues15:10–15:26 Chrome · 23 个标签页 —— MDN、GitHub issues 15:26–15:58 Steam + game session — GPU 82%, 71°C, fan up15:26–15:58 Steam + 游戏 —— GPU 82%、71°C、风扇提速 source: local SQLite file, 214 KB来源：本地 SQLite 文件，214 KB
Sample output — a fresh install starts empty and fills up as you use your machine. Browser tabs and GPU/temperature lines require enabling the browser and GPU monitors in settings (off by default).示例输出——新装的机器一开始是空的，数据随使用逐步积累。浏览器标签页和 GPU/温度两行需要先在设置中启用 browser 和 GPU 监控器（默认关闭）。
Your AI assistant already knows everything on the internet. Kynoptic gives it the one thing it's missing: what actually happened on your machine. Browser pages and GPU/temperature questions need the corresponding monitors enabled in settings — app usage and machine basics are on by default.AI 助手知道整个互联网，却不知道你的电脑上发生过什么。Kynoptic 补上的正是这块空白。浏览器页面、GPU 温度这两类问题需要先在设置中启用对应监控器；应用使用和机器基本指标默认开启。
Someone linked a library last week — Slack, Discord, or HN? You don't have to guess where to search. Your AI looks through your machine's history and tells you.
上周有人发过一个库的链接——Slack、Discord 还是 HN 来着？不用再猜该去哪搜，AI 直接翻这台机器的历史告诉你。
You remember using it Tuesday afternoon — just not the name. Your AI can look it up instead of guessing.
记得周二下午用过，就是想不起名字。AI 直接查记录回答你，而不是瞎猜。
Check what the GPU, thermals, and foreground apps were doing at that exact minute — then the weekly summary writes itself.
查一下那一分钟 GPU 温度、负载和前台应用在干什么；每周的时间总结也能顺手生成。
Install the app, add one block to your MCP client config — your AI can now query your machine's history. Two things to expect: the installer is unsigned, so Windows SmartScreen may show a warning on first launch (choose "More info" → "Run anyway"); and after installing, restart your MCP client (Claude Desktop / Cursor) so it picks up the new PATH before filling in the config.装上应用，在 MCP 客户端配置里加一段——AI 就能查询这台机器的历史了。有两件事提前说明：安装包未签名，首次运行时 Windows SmartScreen 可能弹提示（选"更多信息"→"仍要运行"）；装完后请先重启 MCP 客户端（Claude Desktop / Cursor），让它读到新的 PATH，再填配置。
{
"mcpServers": {
"kynoptic": {
"command": "kynoptic",
"args": ["mcp"]
}
}
}
# current machine status当前机器状态 $ kynoptic now # query yesterday's app/window history查昨天的应用/窗口历史 $ kynoptic query --from yesterday --bucket app/window --json # or query the history as plain SQL — the database lives in the也可以直接用 SQL 查历史，库文件在安装目录的 data\kynoptic.db # install folder under data\kynoptic.db, or wherever KYNOPTIC_DB points（可用环境变量 KYNOPTIC_DB 修改位置） $ sqlite3 "%LOCALAPPDATA%\Programs\Kynoptic\data\kynoptic.db" "SELECT * FROM events" # the DB uses WAL mode: for read-only queries or copying the file,库使用 WAL 模式：只读查询或复制文件时，请连同 -wal / -shm 文件 # keep the -wal and -shm files together and don't truncate them, or you一起保留，不要单独截断它们， # may lose data否则可能丢数据
Kynoptic records events, not screen content: which app was in front, for how long, plus machine metrics. No screenshots, no OCR. By default keyboard data is stored only as per-minute counts and per-key frequency — never the text you typed. A raw per-keystroke recording mode also exists and is off by default; it can be turned on in settings if you explicitly want it. Everything lands in one local SQLite file — the app itself has no upload code path, and the source is open. When you ask your AI a question, the answer is computed on your machine; no third-party LLM is involved unless you plug in your own API key.Kynoptic 记录的是事件，不是屏幕内容：哪个应用在前台、多久、加上机器指标。没有截屏、没有 OCR。默认情况下，键盘数据只按每分钟计数和每个按键的频次存储——不会存你输入的文字。软件里也存在逐键原始记录模式，默认关闭，只有你在设置中明确开启才会记录。一切都在本地一个 SQLite 文件里——应用本身没有上传代码路径，源码公开。你向 AI 提问时，答案在你的机器上算出来；除非你自己填入 API key，否则没有任何第三方大模型参与。
If you've seen screen-recording memory tools before: those rebuild what your screen showed. Kynoptic takes a different trade-off — structured events only, at a fraction of the cost and privacy surface. (Details in the repo's README and BENCHMARKS.md.)如果你见过录屏类记忆工具：它们重建的是"你的屏幕看过什么"。Kynoptic 走了另一条路——只存结构化事件，成本和隐私面都小一个数量级。（详细数据见仓库的 README 和 BENCHMARKS.md）
| Screen recording + OCR录屏 + OCR | Video + LLM summaries录屏 + 大模型摘要 | Kynoptic — structured eventsKynoptic —— 结构化事件 | |
|---|---|---|---|
| What is stored存了什么 | Continuous screen video + OCR index连续屏幕录像 + OCR 索引 | Video chunks + screenshots, summarized by a VLM视频片段 + 截图，交给视觉模型总结 | Event-level records only — no pixels, ever只有事件级记录——完全不碰像素 |
| Running cost运行开销 | High — continuous OCR & indexing高——持续 OCR 与索引 | High — constant VLM API calls高——持续调用视觉模型 API | Near zero — lightweight native capture接近零——轻量原生采集 |
| Disk usage磁盘占用 | GB-scale per week每周 GB 级 | MB–GB scale with auto-cleanupMB–GB 级，依赖自动清理 | KB-scale idle days; tens of MB/day in regular use (see README disk budget)空闲日 KB 级；常规使用约几十 MB/天（详见 README 磁盘预算） |
| AI accessAI 接入 | API + MCPAPI + MCP | Markdown export only仅 Markdown 导出 | MCP server + CLI + open SQL schemaMCP 服务器 + CLI + 开放 SQL schema |
Nowhere. Your data stays in one SQLite file on your machine. There are no servers, no telemetry, and no upload code path. The only outbound network call is a version check against GitHub — first about 2 minutes after the tray starts, then every 24 hours (disable any time in settings). It is a single GET that fetches the full GitHub Releases list (about 100 KB, and it grows with the number of releases and assets); when an update is found, one extra GET downloads SHA256SUMS.txt for verification — static files only, no machine data is ever sent — and it never installs anything on its own (you can verify this in the README).哪也不去。你的数据就在你机器上的一个 SQLite 文件里。没有服务器、没有遥测、没有上传代码路径。唯一的外联是到 GitHub 的版本检查——托盘启动约 2 分钟后首查，其后每 24 小时一次（设置页可随时关闭）。它是一次 GET 拉取 GitHub 完整 Releases 列表（约 100KB，随 release 数和资产数增长）；发现新版本时会多一次 GET 下载 SHA256SUMS.txt 做校验——只下载静态文件，不发送任何本机数据——它绝不会自动安装任何东西（README 里有说明）。
No screens, ever. Kynoptic stores counts and metadata (which app, how many keys), not what you typed. By default keyboard data is only per-minute counts and per-key frequency counts; a raw per-keystroke recording mode also exists and is off by default — it only records if you turn it on in settings. If you'd rather have no keyboard data at all, the whole keyboard monitor can be turned off in settings.永不录屏。Kynoptic 存的是计数和元数据（哪个应用、按了多少键），不是你输入的内容。默认情况下，键盘数据只有每分钟计数和每个按键的频次；软件里也存在逐键原始记录模式，默认关闭——只有你在设置里开启才会记录。如果你完全不想采集键盘数据，也可以在设置里直接关闭整个键盘监控。
Designed for near-zero overhead: no video encoding, no OCR, no cloud calls. Writes are batched to minimize disk I/O; idle days are KB-scale, regular use is roughly tens of MB per day (see the README disk budget).按"接近零开销"设计：没有视频编码、没有 OCR、没有云调用。写入批量落盘以减少 I/O；空闲日为 KB 级，常规使用约每天几十 MB（详见 README 磁盘预算）。
Those rebuild your screen — a replayable record of everything you saw. Kynoptic never captures pixels; it stores structured events: which app, how active, what the hardware was doing. A fitness tracker for your computer, not a black box for your screen.它们重建的是"你的屏幕看过什么"——一份可回放的记录。Kynoptic 从不抓取像素，只存结构化事件：哪个应用、多活跃、硬件在干什么。它是电脑的健身手环，不是屏幕的黑匣子。
Any MCP-capable client: Claude Desktop, Cursor, and other MCP hosts — via the built-in server. No MCP? Use the CLI, or run SQL against the open schema directly.任何支持 MCP 的客户端都行：Claude Desktop、Cursor 等，通过内置服务器接入。不用 MCP 也可以走 CLI，或直接对开放的 schema 跑 SQL。
Windows first — it's where the demand is, and plain Win32 APIs make the monitors truly lightweight. macOS and Linux are on the roadmap. Apache-2.0 licensed, fully auditable.先做 Windows——需求在这里，而且原生 Win32 API 能做到真正的轻量。macOS 和 Linux 在路线图上。Apache-2.0 协议，完全可审计。
Grab the installer from GitHub Releases — SHA256 checksums included. Prefer email? Leave your address and we'll ping you when macOS / Linux land.去 GitHub Releases 拿安装包，附带校验和。也可以留个邮箱，macOS / Linux 版发布时通知你。
Got it — we'll email you when macOS / Linux are ready.收到，macOS / Linux 版发布时发邮件通知你。
Your email is processed by the third-party form service FormSubmit.co solely to send this notification. Prefer to watch the code? Star the repo instead — releases are announced there too.你的邮箱由第三方表单服务 FormSubmit.co 处理，仅用于发送发布通知。更想盯代码？直接 Star 仓库，Release 也会同步发在那里。