Blog

Notes from building
Korely

Engineering decisions. Product trade-offs. Honest updates from the founder.

Does a memory layer still beat naive context? An equal-budget test
AI & Machine Learning

Does a memory layer still beat naive context? An equal-budget test

At the same token budget, does a curated memory block beat just sending the most recent messages? On a 178-question benchmark the right pages answer 76 percent correctly versus 42 percent, and the gap holds even when you swap the reader.

AI & Machine Learning · June 26, 2026 · 10 min
How Korely fuses keyword and meaning into one search result
Architecture

How Korely fuses keyword and meaning into one search result

Plain keyword search misses paraphrases. Plain semantic search misses exact identifiers. Korely runs both and fuses them with Reciprocal Rank Fusion.

Architecture · May 14, 2026 · 9 min
How Korely finds notes by meaning, on your machine
AI & Machine Learning

How Korely finds notes by meaning, on your machine

A small open-weights embedding model running on your CPU turns every note into a 768-number vector. No API key, no upload, no cost per query.

AI & Machine Learning · May 12, 2026 · 8 min
How Korely builds your knowledge graph, on your machine
Architecture

How Korely builds your knowledge graph, on your machine

People, projects, topics, dates. Korely pulls them out of your notes with GLiNER and stitches them into a graph that auto-links related work.

Architecture · May 10, 2026 · 7 min
How Korely turns audio into notes, on your computer
Product

How Korely turns audio into notes, on your computer

Whisper transcribes recordings and audio files on your computer, on every plan, and the audio never leaves it. On Personal and Business the summary is written from the text on servers in Europe.

Product · May 8, 2026 · 8 min
How Korely uses cloud AI: a large model, only when it needs to think
AI & Machine Learning

How Korely uses cloud AI: a large model, only when it needs to think

Local models handle indexing and graph building. Larger cloud models are reserved for chat with citations, post-meeting recaps written in Europe, and the hard summary work.

AI & Machine Learning · May 6, 2026 · 7 min

Frequently asked questions

What is Korely? +

A local-first knowledge base for your meetings, notes, and ideas. It lives on your machine and connects to your AI tools, Claude, Cursor, ChatGPT, via the Model Context Protocol (MCP).

Can I import my Obsidian vault? +

Yes. Korely reads your Obsidian markdown vault as-is, folders, frontmatter, wikilinks, attachments. Notion and Apple Notes import are coming next.

How is this different from Notion or Mem.ai? +

Your notes live as plain markdown files on your disk, kept only there unless you turn on sync. On Personal and Business, a meeting's transcript is also sent to write its recap, and Korely does not keep it. And your AI tools reach the vault directly via MCP, not through a proprietary chat box you don't control.

Is it open source? +

Not yet. The desktop app and MCP server may go open source down the line. Building in public means showing the work; the source comes when it's stable enough not to mislead anyone.

How much does it cost? +

Free forever for the local vault, unlimited notes, hybrid search, MCP, Obsidian import. €14/month for Business: live meeting transcription, audio/video to note, AI chat with GraphRAG over your vault, cloud sync, MCP cloud multi-tenant.

Ready to try it yourself?

Free forever. Your notes stay on your computer until you turn on sync.