Context Passport
Your context, owned by you, handed to each AI tool in slices.
The itch
Every AI tool meets you as a stranger. The planner knows how you think about roadmaps, the editor knows your conventions, the drafting tool knows you hate the word “leverage”, and none of them talk to each other. Vendors are fixing this by remembering you inside their own walls. The better shape is a passport: one document you own, stamped by the tools that have read it, handed over in slices. It keeps itself current as you work, so there’s no queue at the passport office to renew it.
How it works
Design decisions
The store
| Option | Good at | Weak at |
|---|---|---|
| Markdown files in git | Readable, portable, every change has a history | Search beyond keywords |
| SQLite (full-text plus a vector extension) | One local file, fast search by keyword and by meaning | Not something you’d open and read |
| Vector database | Similarity search at scale | Overkill for one person; facts become opaque chunks |
| Graph | Relationships between people, projects and facts | Heavy to model and keep tidy |
The pick: Markdown is the source of truth, one file per page. SQLite is an index rebuilt from those files, so it can be thrown away and regenerated any time. Retrieval checks scope first and similarity second, so a page an agent isn’t allowed never reaches the search at all.
Intake
| Decision | The call | Priority |
|---|---|---|
| Auth | OAuth per source, read-only scopes, tokens kept on the device | Must |
| Push or pull | Agents push at the end of a session over MCP; docs and email get pulled | Must |
| How often to pull | Once a day in a batch | Must |
| Worth keeping? | A small model separates lasting facts from passing chatter, and raw text is dropped once checked | Must |
| Already recorded? | Match by meaning against the page. Same fact refreshes its date; a changed fact goes to review, with the old version kept | Must |
| Webhooks instead of polling | Gmail push and Notion webhooks, so changes arrive within minutes | Should |
| Hand-written notes | A photo or screenshot goes through a vision model, then the same filter | Should |
| Auto-accept | Skip review above a confidence threshold, so the queue stays short | Should |
| Contradictions across sources | Flag when two sources disagree about the same fact | Could |
| Real-time capture | Every message, as it happens | Won’t |
| Keeping raw documents | Only extracted facts are stored, never the email or doc itself | Won’t |
| Writing back to sources | The passport reads from your tools; it never edits them | Won’t |
Under the hood
- MCP as the handover. The passport runs as an MCP server, so any tool that speaks the protocol asks for context instead of being pasted into.
- A small model for intake. Filtering and matching are cheap, narrow jobs, which makes them a good test of how far a small, fast model goes before a frontier one is needed.
Where it goes
The personal version is a concept. The bigger version of the same idea is being built for the enterprise: an organisation’s AI context, scattered across tools and teams, brought together into one shared knowledge layer.