Inspired by David Rockefeller's card file: remember the people you meet and the moments you share. Thought Pins brings that habit into a searchable personal record, with memories linked to their sources and exports you can keep.
Open-source journaling with structured recall and portable exports.
Website · Run locally · Codebase walkthrough · Documentation
I originally built Thought Pins for my own use. I decided to open-source it
in case anyone else found it useful and wanted to use or adapt it.
A personal record you can return to
Thought Pins turns journal notes, conversations, voice notes, and saved reading
into a connected record of people, places, events, and ideas. Ask about a past
moment, inspect its source, or export your own record as plain files.
The product includes a responsive React web app, a native SwiftUI iPhone/iPad
app, a FastAPI backend, and an Obsidian-compatible vault. Five destinations —
Recap, People, Chat, Places, and Pins — keep the experience consistent across
screens. Android has a separate client and build checks.
Product capture uses a fictional account. Mobile and desktop previews.
The project is in active development. Source, automated checks, and setup
instructions are public; hosted availability and store distribution are
separate release decisions. See validation and limits
for what the evidence supports.
Capture in the web app
| Input | Current behavior and limits |
|---|
| Voice note | Record, listen back, choose journal or reference source, then explicitly save or discard. A failed save keeps the in-memory recording for retry with the same request identity. Leaving the view or refreshing discards an unsaved recording. |
| Phone photo or screenshot | Apply the photo's orientation before local OCR. Printed text works best; handwriting, equations, and page layout fidelity are not established. |
| PDF and text notes | Extract embedded PDF text from up to 100 pages. Scanned pages without text are read with local OCR (Tesseract) by a background worker job, up to 100 pages and 8 minutes per PDF: the upload is saved and answered at once, and the source is ready for recall when its pages have been read (a server without background processing reads up to 20 pages within 40 seconds during the upload instead); embedded text is never replaced by OCR, turned pages are read upright, and pages stored as strips are reassembled. Page limits, OCR'd pages and pages still unread are reported with the reason. Handwriting and page layout fidelity are not established. Files are limited to 25 MB. |
| Word and PowerPoint | Read .docx paragraphs, tables, footnotes and endnotes, and .pptx slides in presentation order with speaker notes, each labelled by slide number. Tracked deletions are excluded. Text inside images, legacy .doc/.ppt, and password-protected files are not read; the upload says which. |
| Article link | Save readable authorized content when available. A saved link is distinct from extracted text; the app does not bypass access controls. |
Uploads distinguish unreadable, partial, queued, and completed extraction.
Attachments in Chat default to reference sources, with an explicit journal
choice; new recordings default to the personal journal. Source browsing can
page past the first 100 items and search titles, authors, and sites across the
user's library. Incoming replies preserve the reader's position in chat history.
The next product evidence should come from observed web capture/recall tasks
with 5–10 consenting users and a separately scored end-to-end answer evaluation.
Automated workflow checks and retrieval scores do not establish adoption,
retention, transcription accuracy, or answer accuracy. Further native expansion
and ranking complexity should follow measured failures in those workflows.
Alternatives, briefly
Thought Pins groups journal capture, people and place views, source retrieval,
and export in one application. The tools below overlap with different parts of
that workflow.
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
