Lapse

The app is the failure point

Every note-taking tool asks the same thing: stop what you're doing, open the app, type. That stop is where it breaks. The thought worth keeping usually shows up while you're doing something else entirely, and by the time Notion's open it's already gone.

So I built the opposite. A journal you write by talking, where the AI you're already mid-conversation with is the pen, and a private database you own is the paper. You say "log this," it's saved in your own words, and nothing else about how you work changes.

Sign up, connect, talk

Three steps. Sign up with email and password, your journal scoped to you. Connect your AI, OAuth for Claude, a token for everything else, local models included, fully offline if you want that. Then just talk. "Log this: ..." in any chat, and it's done.

Because the AI can read its own past entries back, this isn't a dead archive. Ask it what you shipped last week and it pulls the week and reflects the pattern back at you. It remembers what you told it, which is the part a plain notes app never does.

Runs on Cloudflare Workers and D1. Free to run. Passwords and tokens never stored in plaintext, entries never shared.

Every agent writes to the same log

It's just an MCP endpoint, so it doesn't care which agent is holding the pen. Claude in the browser, Claude Code in a terminal, opencode, a local model running fully offline, point any of them at the same URL and they're all writing into the same journal.

That matters more than it sounds. Claude Code can log why it made an architecture call at 1am, and the next morning Claude in chat reads that entry back before you've said a word about it. The knowledge stops being trapped inside whichever tool happened to be open when you had the thought. One canvas, every agent adding to it, nobody starting from zero.

A two-year-old notebook entry, finally buildable

I wrote this idea down two years ago: a system where I don't think twice before speaking, and something takes care of remembering for me. At the time it was impossible, no model could reach out and use a tool on your behalf mid-conversation.

MCP changed that math. The bet in that notebook was that LLMs would eventually call external tools on their own, and only the ideas that needed exactly that would become buildable. This is one of them.

This is live and I use it daily, but it's early. Rough edges I already know about: onboarding could be smoother, the read-back experience needs more shape, and I haven't stress-tested it against a client that isn't Claude yet.

Feedback is genuinely welcome, especially from anyone who's tried to make an AI remember things across sessions and hit the same walls.

lapse.in