Rocketnotes

fynn-fluegge‘s “web-based Markdown note taking app with native AI feature integrations like chat, text completion, voice-to-text transcriptions and agentic document archiving.” Apache-2.0, 1.4k★ / 77 forks, Go + TypeScript + Python. T1 as an official project repository, and everything below is the project describing itself; nothing here is measured.

It matters to this wiki because of one feature. The llm-wiki thesis is that the memex failed on maintenance and the LLM supplies it. Rocketnotes ships maintenance as a bullet point:

“Zettelkasten with Agentic Archiving: An AI agent analyzes snippets from your ‘inbox’ and intelligently files them into the most relevant existing document.”

That is agentic-archiving, and it is the first time this corpus has seen the pattern’s maintenance half sold as a product feature rather than described in an essay (llm-wiki-gist) or built by one team for itself (gbrain).

What it is made of

• AI: Langchain and Langgraph; multi-provider — “OpenAI, Anthropic and Together AI models” — with Ollama for the fully local mode. Vector store is S3 Vectors in cloud, ChromaDB locally; embeddings via sentence-transformers.
• Store: DynamoDB for documents, S3 for blobs. Notes are markdown content, but the persistence layer is a database rather than a folder of files.
• Shape: Angular web app plus an Electron desktop build, Go/Python Lambda handlers on AWS, or the whole thing in Docker — “Run the entire application on your own machine using Ollama for 100% local and private AI processing.”
• Surfaces: hierarchical document tree with drag-and-drop, content search, semantic search, document sharing, KaTeX and Mermaid, a Neovim plugin, and an MCP server so the knowledge base is callable from any LLM application (model-context-protocol).

Both sides of the RAG argument, in one product

This wiki has treated retrieval and maintenance as competing answers: llm-wiki-gist positions the compounding wiki explicitly against retrieval-augmented-generation — don’t retrieve over raw documents, maintain a distilled artifact instead. Rocketnotes ships both without noticing a tension. “Chat with Your Documents” and semantic search are RAG over your notes; agentic archiving is the maintenance move. Whether they compose or merely coexist is exactly what nobody here has measured.

The name it borrows, and the principle it inverts

“Zettelkasten” is doing work in the marketing that the feature does not do. zettelkasten‘s central claim is connectivity over categorization — “connection, not a collection,” placement doesn’t matter as long as you can link, structure emerges bottom-up from explicit reasoned links between atomic notes.

Rocketnotes’ agent files a snippet into the most relevant existing document, inside a hierarchical tree you reorganise by dragging. That is placement, and a good filing decision is precisely what Luhmann’s method says is not where the value lives. The inbox half is faithful — fleeting notes captured fast, processed later, is real slip-box practice — but the processing step here ends in a folder rather than in a link.

Worth stating plainly because it is a pattern this wiki should expect to keep seeing: the vocabulary of the lineage travels faster than its mechanics. Nothing stops an agent from proposing links instead of destinations; this one wasn’t built to.

Storage, and what the lineage usually asks for

obsidian is valued in this wiki for local markdown files you own, and yank-note states the same bet explicitly. Rocketnotes keeps markdown as the content format but puts persistence in DynamoDB and S3. The local Docker mode keeps the whole stack on your machine, which answers the privacy question; whether it also leaves you a directory of .md files you could walk away with is not something the README says, and I have not read the code.

Related

agentic-archiving · zettelkasten · llm-wiki · gbrain · retrieval-augmented-generation · obsidian · yank-note · model-context-protocol · fynn-fluegge · synthesis