for builders
Plan before you prompt.
The model can do the work. The one thing it cannot do is know what you know.
The model can write the code, draft the email and run the loop. The one thing it cannot do is know what you know.
AI engineering moved through raw prompting, then prompt engineering, then context engineering, and now loops that run on their own. Every stage made the model cheaper to run and the input more valuable. What you feed an agent now determines what it costs you: in money, in visibility, in consequences.
So the scarce resource is not intelligence any more. It is the accumulated context that makes intelligence useful: what was decided and why, who owns what, the reasoning nobody wrote down because writing it down was always the expensive part. That part just got cheap, and most people have not noticed yet.
A company forgets faster than it learns.
The explanation on the whiteboard, the reason behind a decision, the thing a senior person knows and has never written down: all of it lives in conversations, and conversations evaporate. When someone leaves, it leaves with them. When you ask why six months later, nobody is sure.
Preserving that is not a note taking problem, it is an infrastructure one. It has to be captured at the moment it is said, filed without anybody volunteering to file it, and stored in a format that outlives the tool that captured it. Plain Markdown in a folder you own satisfies the last part in a way no database can, because it does not depend on us still existing.
It is just files.
Every note is a Markdown file with YAML front matter. Topics are computed from that front matter, which is why a note can belong to two of them. Transcripts are plain text. Chats are one line delimited JSON file per thread.
There is no proprietary format to reverse engineer, and no API key to obtain before you can read your own material.
---
title: Q3 numbers wait on Priya's pricing sheet
created: 2026-09-04T14:21:07
tags: [pricing, product]
---
The numbers are ready but not sent. Priya owes
the pricing sheet and no date is set.
## Decisions
- Q3 numbers ready, not yet sent to the client
## Open threads
- No date set for Priya's pricing sheetPoint an agent at it today.
Nothing here needs us to ship anything first. These all work with the app as it is.
Export the library
A zip of everything, in open formats, from the all notes screen. Unzip it wherever your agent already reads files.
Use the iCloud copy
With backup on, the library lands in your own iCloud Drive once a day. Point a desktop tool at that folder and your context is there.
Write back into it
Drop a new .md file into the
folder, or edit one in place. The app reconciles on its next scan, and your
edits win.
A memory connector, built the careful way.
The goal is that any assistant you talk to can draw on your own memory. The constraint is that this must not quietly become the content database we deliberately do not have.
Here is the shape it will take, so you can hold us to it.
Off by default
A feature you switch on, per category. Switching it off deletes everything it put on our side, immediately.
Read only, to start
Four tools: your profile, a search over what Vasini knows, one subject at a time, and your open commitments.
Distilled, never raw
Facts and entities, not episodes. Quotes lifted from your transcripts are stripped out; the tap through to the source stays on your phone.
Every read is logged
A screen in the app that says which assistant read what, and when.
What it never carries
Your transcripts, recordings or notes in full, sitting on our servers so somebody else’s model can read them.
There is no npm package and no API key yet. When there is, this page will carry the real configuration and the real list of what it exposes. Not before.
