The architecture question
Every notes app now claims AI. The claims hide three very different architectures, and choosing between them is the real content of this query. In the first, the vendor's cloud model reads your notes on the vendor's servers (Notion AI, Mem). In the second, the app is a front end and you bring the model relationship, through connectors to Claude or ChatGPT (Capacities). In the third, your notes stay local and an AI reaches them through a bridge you run, typically an MCP server, from your own machine (Obsidian with plugins, AFFiNE, Joplin). Privacy, cost shape, and capability all follow from which architecture you pick, so this guide grades architecture first and features second.
Obsidian itself belongs to the third architecture by construction. There is no official Obsidian AI. What exists is an ecosystem: community plugins that call the model APIs you configure, local-model options for the fully private route, and MCP bridges that expose your vault to an assistant running on your machine. The strength is sovereignty: your vault is files, your keys are yours, and no vendor can reprice your own notes back to you. The weakness is assembly: you are the integrator, and the experience is exactly as coherent as the plugins you chose.
The field, by architecture
| Tool | Where the model runs | What the AI can see | The trade |
|---|---|---|---|
| Notion AI | Vendor cloud | Your workspace, as you permit | Most integrated, least sovereign |
| Mem | Vendor cloud | Everything you capture into it | AI organization instead of folders |
| Capacities | Your Claude or ChatGPT account via connectors | The objects you connect | Structured notes meet frontier models |
| AFFiNE | Your assistant via its MCP server | The workspace you expose | Local-first with a bridge you control |
| Joplin | Your assistant via its MCP server | The notebooks you expose | Open source notes with an AI door you open |
| Obsidian (baseline) | Your keys, your plugins, your machine | Your vault, as configured | Maximum control, maximum assembly |
Notion AI is the answer when the team wants AI as a feature of the workspace rather than a project: summarizing, drafting, and querying across pages, with the enterprise controls a company expects, at the price of your notes living and being processed in Notion's cloud. Mem is the capture-first answer: throw everything in, let AI organization replace filing, and accept the same cloud trade in a more opinionated shape. Capacities is the interesting middle: a structured, object-based notes system that connects to Claude and ChatGPT, so the model relationship is yours while the workspace stays designed rather than assembled. AFFiNE and Joplin are the bridge answers for readers who like this query's premise but want the assembly done: local-first or self-hosted notes with an MCP server as the AI's doorway, so an assistant can search and draft against your notes under your roof.
One boundary deserves stating because marketing blurs it: an MCP bridge on your machine exposes your notes to the assistant you connect, under that assistant's terms. "Local notes" plus "cloud model" means the excerpts the model reads travel to the model. The genuinely private route is a local model, which Obsidian's plugin ecosystem and the bridge tools can both accommodate, at a real capability discount. Choose with that sentence in view.
