Repository navigation
Harness.WinUI 1.1.0
Long-term memory (Settings → Memory, off by default). The agent can save short notes about you and your work, find them again, update or forget them. Every note is approved by you first, and nothing that looks like a key or a password is ever saved. Notes are plain Markdown files in a memory folder that you can read and edit. At the start of a conversation the most relevant ones are put in front of the model, chosen once so the prompt cache keeps working. Notes unused for 90 days stop being used until you bring them back.
Memory window. Your memories as a list, a timeline and a connection graph, with search over memories and past conversations. Click a memory to edit it, forget it (with undo) or open the conversation it came from. Under a conversation's first reply, "Used N memories" shows which ones the agent was given.
Tidy up. Your model looks at the notes and recent conversations and suggests what to add, merge, rewrite, drop or confirm. You see how much will be sent first and pick which suggestions to apply; replaced and dropped notes are kept so they can be restored.
Search past conversations (Settings → General, off by default). The agent can look through your saved conversations to answer "when did I work on this?". Each search asks for your approval, and the answer links to the message.
Download
| File | For |
|---|---|
Harness.WinUI-1.1.0-win-x64.zip |
Most Intel / AMD PCs |
Harness.WinUI-1.1.0-win-arm64.zip |
ARM PCs (e.g. Snapdragon Copilot+ PCs) |
Unzip to any folder and run Harness.WinUI.exe. Requires Windows 10 1809+ and the WebView2 Runtime (built into Windows 11). The executable is not code-signed, so SmartScreen may warn the first time: choose More info → Run anyway. Verify downloads with SHA256SUMS.txt.
長期記憶(設定 → 記憶,預設關閉)。agent 可以把關於你和你工作的短筆記存下來、找出來、更新或忘記。每則都先經你核准,看起來像金鑰或密碼的內容一律不會存。筆記是放在 memory 資料夾裡的 Markdown 檔,你可以直接閱讀、編輯。對話開始時,會把最相關的幾則放到模型面前(只選一次,所以提示快取仍然有效)。超過 90 天沒用到的筆記會暫停使用,直到你把它找回來。
記憶視窗。 用清單、時間軸、關聯圖檢視你的記憶,搜尋涵蓋記憶與過去的對話。點一則記憶可以編輯、忘記(可復原)或打開它來自的對話。對話第一則回覆下面的「用到 N 則記憶」會告訴你 agent 拿到了哪些。
整理記憶。 你的模型看過筆記與最近的對話後,建議哪些該新增、合併、改寫、刪除或確認。送出前會先告訴你要送多少內容,由你挑要套用哪些;被取代或刪除的筆記會保留,可以還原。
搜尋過去的對話(設定 → 一般,預設關閉)。agent 可以翻你存下來的對話,回答「我是什麼時候做這件事的?」。每次搜尋都先經你核准,答案會附上連到那則訊息的連結。
下載後解壓縮,執行 Harness.WinUI.exe 即可。