Local first. Clear about what leaves.
Memory is primarily local. Cloud models receive relevant text and memories. Optional OCR does not save or upload screenshot originals, but recognized text may enter model requests.
Your local-first personal desktop assistant
Pause for a Cue. Add / to ask. Your memory lends a hand, while the main app stays out of your way.
Real overlay UI · Public samples, no desktop or model access. Take over anytime. App UI in English.
The design you reviewed this morning. A decision from the afternoon meeting. A note before you leave. The information is scattered; your life is connected. HaroCue starts with one personal space where these clues can become useful memory.
Enable recording and grant permission, then keep reading, writing and working. HaroCue captures text from allowed foreground windows, uses your configured model to extract clues, and connects events, decisions and sources into memory. When you ask later, that context comes with the answer. Follow one work note through the whole process.
Try the controls. Public samples only: no desktop access or model connection. Refresh to reset. The app’s current UI is shown in English.
The four steps play automatically. Select any step, open a memory in the graph, and trace it back to the original work note.
Explore the memory UI ↗Forgot a detail midway through writing? Cue offers an evidence-backed next sentence when you pause. For an explicit question, add /. Review the answer, tool activity and sources, then press Tab to accept.
Stay in your original input. Clear it, and the old suggestion leaves too.
Availability depends on accessible app text and system permissions. Unreadable content is not invented.
Review by day, week or month. Start with sourced entries, then let the assistant turn them into prose. Distinguish what you did, what you read and what remains unconfirmed. Check before copying; you decide when and with whom to share.
Keep using Codex, Cursor or Claude Code. Explicitly call HaroCue through MCP to pass relevant local context and continue from the progress you already made.
Recording and desktop assistance have separate switches. Pause new recordings and still ask an explicit question. Connect your own cloud model or use local Ollama.
Memory is primarily local. Cloud models receive relevant text and memories. Optional OCR does not save or upload screenshot originals, but recognized text may enter model requests.
Sources, corrections, stop-use and forgetting are available alongside memory. You can decide what remains useful.
Recent desktop features are in local trials on Apple Silicon macOS. Public packages, available features and permissions are stated separately on Downloads.
Check available builds ↗The local core is free. Plus and Pro managed services are planned, with proposed prices and benefits labelled clearly. In-app estimates are not provider invoices.
Compare plans ↗Product changes go into the blog and issues have a public home. Data handling, cancellation, refunds and terms are available before you start.
Policies and support ↗User feedback
“I no longer reopen three or four pages just to write a confirmation email. The dates, reasons and open questions come back together.”

“The constraints we agreed in a meeting are still there when I return to the document. It reminds me of the missing sentence without rewriting the whole paragraph.”
“Add / at the end of a question and the answer comes back where I am writing. No copying and pasting or opening another AI window.”
“When I organize interviews, names and constraints from the previous page follow me into the research document. I can focus on interpreting the material.”
“It stops at password fields, and pause and delete are easy to find. Knowing what will not be remembered makes me comfortable keeping it nearby.”
“Changes from the meeting carry through to delivery documents. I keep working in Cursor, and HaroCue connects the context.”
Independent developer of HaroCue
I built HaroCue because coding Agents are increasingly capable, but still cannot see what you just did outside the conversation. I want recent work context to stay local and return only when it is genuinely useful.
I studied Computer Science and worked on reliability-sensitive embedded software. I currently design, build, and maintain HaroCue across context capture, local memory, MCP access, and cross-platform releases.
HaroCue began as the open-source DayRecord project. Its releases, issue handling, and product changes can all be verified publicly.
View the public repositoryNo. Start in the personal life-and-work space. Create a separate project space when a particular context needs a boundary.
No. Automatically observed content can be labelled as an assistant inference, distinct from confirmed information. Check sources, then correct, stop using or forget it.
Yes. Recording and desktop assistance have separate controls. Pausing stops subsequent persistent capture and automatic processing; in-flight requests may finish. Explicit questions still require the relevant permissions and context.
Memory is primarily local. Cloud models and online tools receive relevant content needed for a request. Optional OCR does not save or upload screenshot originals; recognized text may enter model requests.
The demos use the current app’s UI components, displayed in English. Try questions, graph and list views, corrections, reviews, cursor assistance and independent controls. Everything uses public samples without a live model; refresh to reset.
Start with one small thing you want HaroCue to remember.
Download