Show HN: Reor – An AI note-taking app that runs models locally

Original Article ↗ Hacker News Discussion ↗

An open-source project called Reor aims to turn local Markdown note collections—especially Obsidian-style vaults—into an AI-augmented knowledge base by running language models entirely on the user’s machine. Commenters explore use cases like semantic search, related-note discovery, and Q&A over long-term personal notes, while weighing concerns about over-reliance on AI for thinking, data ownership, and interoperability. Many praise the privacy and flexibility of plain-text workflows but report mixed results and technical issues with local models, and express interest in features such as multiple vaults, plugins, mobile support, and better handling of formats like PDFs and front matter.

Project overview and intent

  • Reor is a local-first, open-source AI note-taking app operating on Markdown “vaults.”
  • Emphasis on privacy, offline use, and interoperability with other tools that use plain files.
  • Several commenters see it as a prototype of how desktop software may integrate local LLMs.

Integration with existing tools and formats

  • Works on directories of Markdown files; plain text support is planned.
  • Can be pointed at existing Obsidian vaults; operates 1:1 on the filesystem and syncs changes via a vector DB.
  • Users request:
    • Direct Obsidian plugin instead of separate app.
    • Multiple independent vaults / context switching (already in PR).
    • Better handling of YAML front matter and Logseq-style outlines (current behavior can mangle these).
    • Importers for OneNote, PDFs, browser bookmarks/history.

AI’s role in knowledge management

  • Debate over whether AI-assisted organization improves or harms thinking:
    • Concern that offloading organization to AI may weaken active thinking and personal mental models.
    • Counterpoint: AI can surface inspiration and connections; best used selectively, not as a crutch.
  • Some argue personal knowledge graphs should be organized by humans, with LLMs used for discovery and queries.

Model quality, prompts, and RAG

  • Mixed experiences with local models:
    • Q&A often helpful for summaries and overviews but unreliable on specifics; related-notes links can be weak.
    • Smaller 7B models often feel “too dumb”; larger models (34B–70B, Mixtral) work better but are heavy.
  • Discussion that chunking isn’t the main issue; prompt design and tuning are critical but underemphasized.
  • Suggestions to use lightweight models fine-tuned for summarization, and explore vector/graph DBs and neuro-symbolic approaches.

Performance, hardware, and stability

  • Reports of crashes on Linux, UI lockups on M1 Macs, and errors when vaults are large or context windows overflow.
  • Some users have good results on M1/M2 Macs and RTX GPUs; others find CPU-only too slow or unstable.
  • Calls for clearer hardware requirements and better GPU utilization.

Data ownership and storage model

  • Strong support for using plain Markdown files over app-owned databases to avoid lock-in and enable multiple tools.
  • Some argue databases can also be shared, but filesystem-as-database is seen as simpler and more universal.