Why Your Second Brain Shouldn't Live in the Cloud.
The category is now well-established. Notion, Obsidian, Mem, Reflect, Saga, Tana, and a dozen other tools all promise the same thing: a personal knowledge system that captures…
The category is now well-established. Notion, Obsidian, Mem, Reflect, Saga, Tana, and a dozen other tools all promise the same thing: a personal knowledge system that captures what you know, what you have read, and what you want to do — and then surfaces it intelligently when you need it. This is the "second brain" thesis, and the products selling it have raised hundreds of millions of dollars on the premise.
There is a problem with that promise. The brain you are paying for lives in someone else's data center.
The Bargain Most People Don't See
When you pay $10 a month for a cloud second brain, you are not just paying for storage. You are paying for the convenience of letting the company own your reading list, your calendar, your medical journal, your relationship notes, your draft messages, your goals, your health metrics, and the embedded reasoning their model has built up about who you are.
That data is encrypted in transit. Often encrypted at rest. Their privacy policy says it is not used to train models. You believe them, mostly.
And then one of three things happens.
The company gets acquired. The privacy policy changes. The new owner has a different definition of what training data is. Your archive is renegotiated.
The company runs out of money. Their data is sold as part of bankruptcy proceedings. Buyers of distressed startup data are not buying it for sentimental reasons.
The company is breached. Six years of your private notes show up on a data dump site, indexed by a search engine, queryable by anyone who knows your email.
These are not hypotheticals. Each has happened in the last twenty-four months to a major productivity tool. The only thing that varied was which one. The pattern is not going to stop.
The Local-First Alternative Is Older Than You Think
Local-first is not a 2026 idea. The entire history of personal computing was local-first. You owned your files. They lived on your machine. Software ran on your CPU. The cloud became dominant because syncing was hard and storage was expensive — both of which are now solved problems.
What changed in the last few years is that the AI models are also small enough to be useful on local hardware, or — when you do need a frontier model — small enough in the prompt that the call can be scoped, transient, and not training data.
This is the design premise behind Elyra. Your profile, your goals, your tasks, your morning briefs, and your reasoning history live as plain markdown and SQLite on your laptop. The Whoop data syncs to local storage. The calendar syncs to local storage. When the agent needs to reason, it sends a focused prompt to Claude with just the slice of context relevant to today — and nothing else. There is no rolling cloud index of your life.
What "Local-First" Actually Costs You
It is fair to ask what you give up.
Cross-device sync is harder. If your second brain is on your laptop and you are away from the laptop, you do not have the brain. The honest answer is that you replace cloud sync with iCloud Drive, Dropbox, or a self-hosted Syncthing folder. None of them are as integrated as Notion's first-party sync. They are also not selling your data.
Search across your entire archive is slower. A cloud product can build expensive indexes server-side. A local product builds them on your machine. For most archives this is fine. For a 200,000-note archive, it is a real engineering problem.
You manage your own backups. This is a feature, not a bug. The cloud backup is also a single point of failure. Most people who lose their cloud archive lose it not because the company died — they lose it because they did not pay for the export feature on the way out.
These are real trade-offs. They are also bounded, solvable, and shrinking each year as local hardware gets faster and tooling matures.
The Specific Risk That Drove This Decision
When we started building Elyra, the design constraint was simple: assume the worst-case data exposure for a user. What is in this archive that you would not want public if the company that built the archive disappeared tomorrow?
For our test users, the answer included: their Whoop recovery scores, their calendar with private events, their reading list (a surprisingly intimate document), their goals for the year, their medical follow-ups, and the agent's internal model of who they are and what they want.
A cloud product builds embeddings and indexes from that data. The embeddings are a dense, lossy summary of the data — which is to say, they are the data. They live wherever the company chooses to put them. They are a single subpoena, breach, or pivot away from being someone else's.
We did not feel comfortable shipping that surface. The local-first design is not an aesthetic choice. It is a threat model.
What "Privately Use Claude" Actually Means
Elyra calls Claude. We do not pretend it does not. The distinction is what gets sent.
A cloud second brain that uses an LLM typically runs continuous embedding generation across your entire archive. Your data is, in a meaningful sense, resident in their pipeline.
Elyra calls Claude only with the focused prompt for the current turn — your morning brief request, your specific question, your single voice command — plus the markdown slice that turn requires. Claude responds. The response is written to local markdown. The transient prompt is gone.
There is no background indexing of your life. There is no embedding of your private goals into some vector store. The cloud model is a function call, not a tenant.
This is what "local-first" should mean in 2026: not "we never use the cloud," but "the cloud never holds our state."
The Test To Run On Any Personal AI Tool
Before you put a year of your life into a personal AI tool, ask the company three questions in writing.
Where does my data live? If the answer involves anything other than "your device," follow up with: who else has read access?
What do you embed? If the company runs continuous embedding across your archive, those embeddings are the data, and they live wherever the company puts them.
What happens to my archive if you go bankrupt or are acquired? The honest answer is that the contract you signed will be renegotiated. The local archive cannot be renegotiated, because it is not theirs to renegotiate.
If the answers do not satisfy a paranoid friend who works in security, the answer is local-first.