Karsten Lehmann 18 November 2026 09:00:00
Two days ago we published MindooDB Haven in 2026, the year in four minutes, a short film about what happened on the platform this year. That film is about what Haven can do. Many developers asked us the next question right away: how does it work underneath?
So we made two more videos, in the same style and with the same narrator, and they are now on our YouTube channel. Each takes about five minutes. They are meant as an overview for developers: what the MindooDB platform does for you, how it stores data, and how it prepares that data so that you can search it and sync it quickly. Afterwards, the documentation on GitHub has the details.
Part 1: How MindooDB stores your data
The first video is about the data model, and its message fits in one sentence: nothing is ever overwritten.
- CRDTs and Automerge. Every edit is recorded as a small change, and changes made in parallel, even after a week offline, merge automatically and in the same way on every device. App developers never touch Automerge directly. They work with plain JSON documents.
- Chained, signed, encrypted, appended. MindooDB chains each change to the ones it builds on, signs it with the author’s key, encrypts it on the device and appends it to the store. The server only ever sees ciphertext.
- Time travel. Because the whole history is kept, any document, or a whole database, can be opened as it was at any moment. We wrote about this earlier in application time travel.
- The history as a graph. Haven’s database browser shows the history as a DAG: who changed what, where two people worked in parallel and where their work merged again. That makes it a complete audit trail, valuable when many people work in parallel and just as valuable when an AI agent changes data on your behalf, because every change can be rolled back.
- Delete, undelete and governance. A delete is just one more entry in the chain. The rules about who may do what, the governance, live in the same signed, append-only history.
- Witness time and two tiers of checks. The server stamps every change with a signed time without being able to read it. Rights are checked during sync and again on every device, so a manipulated client can write what it likes locally, but its changes never become part of the shared data.
- Purge. For data that should not live forever, such as free and busy times, a governed purge removes a document together with its whole history.
Part 2: How MindooDB stays fast
If a database never overwrites anything, it collects thousands of changes. Rebuilding every document from its history each time you look would be far too slow. The second video shows how we do that work once, at the right moment, and keep the results.
- Work while you wait. During sync the device mostly waits for the network anyway. MindooDB uses that time to apply incoming changes right away.
- A changes feed. Every changed document lands in an index ordered by the time it changed, much like the changes feed in CouchDB. It works on metadata only, and app developers can use it too, including deletions. It is the easiest way to keep external indexes and databases up to date.
- The document summary buffer. For each changed document, MindooDB extracts just the fields that search and views need, configured per database in a document called
dbsetup. Because the rules live there and not in the documents, you can change what the summary holds at any time without touching a single document. In HCL/IBM Notes/Domino, which values are summary values is part of the data itself. - Virtual Views and direct queries. Sorted, categorized views across one or more databases with totals and drill-down, kept current from the changes feed so that only the entries that changed move. Or query the summary buffer directly: filter, sort, find.
- Full-text search and links. An incremental full-text index on MiniSearch that you can combine with field filters (search over encrypted data), and documents that follow their links like a join.
- Local first, but not local only. Data is read locally whenever possible, but the same access works client to server and peer to peer over Iroh. In Mindoo Chat, colleagues can open a file while it is still uploading.
- Smart sync and attachments in chunks. Bloom filters and saved cursors so each side only checks what is new. Attachments you can keep all, none or pin, stored in chunks so that transfers resume, apps can read any part of a file, and growing files such as logs only append new chunks.
Why this matters for developers
None of this is visible to the people using Haven. They simply search and filter in data that is always current. And app developers rarely need to think about it either: the SDK picks the fast path, and our App Builder follows the best practices guide from the MindooDB project. But if you want to build on MindooDB yourself, from Node.js, the browser, React Native or the JVM, ten minutes of video are a good start before you dive into the docs.
Try it
- Watch both videos and subscribe to our YouTube channel. Each video has chapters, so you can jump straight to the part you care about.
- Read the MindooDB documentation. MindooDB is open source under the Apache 2.0 license on GitHub.
- Open Haven in any modern browser at haven.mindoodb.com, or download Haven Desktop for macOS and Windows.