See it in action

Forty-five seconds of Memento doing its job, then five short clips, one per feature. All of it is the real app, recorded against a library the agent built from four captures. Sound off, captions on.

Capture from a phone, the agent files it, ask your library

The whole loop: a note shared from a phone, filed onto the right pages while the laptop stays closed, every claim citing the capture it came from, a bigger edit waiting for approval as a diff, and a question answered with receipts. The one thing cut for time is the wait — filing a capture takes the agent a few minutes, not a few seconds.

Ask for a change

Edit mode — the new fact lands marked as inferred

Switch the assistant from Read only to Edit and say what you want. Small, additive changes land straight away; anything that rewrites a confirmed fact waits in your Inbox as a diff. Here the request was to record a preference someone assumed, and the page gains the line with an inferred label, so it can never be mistaken for something Priya actually said. How the split is decided is on The assistant; the labels themselves are explained on Sources & trust.

Routines

Settings → Routines — a prompt on a schedule

A routine is a prompt written once and run on a schedule, exactly as if you had typed it into the assistant: a morning brief on weekdays, a Friday sweep of open action items. Each routine keeps its own conversation, so it remembers what it told you yesterday. Details on Routines.

Planner

Planner — every action item, from every page, in one place

Action items live on the pages that produced them, where the context is. The Planner is the same lines gathered across all your spaces: tick one off here and the page changes too, open one and you land on the page. Filing extracts commitments from captures automatically. More on Planner.

Locked down, never locked in

Network rules, then a page's download menu

The agent works in a sandbox that can reach nothing you have not approved: each new host it wants waits for your decision, and the platform hosts it needs are listed in the open. In the other direction, every page is a plain Markdown file in a Git repository, downloadable one at a time or as the whole library with its history. See Spaces & tuning for the network settings and Library & editor for export.

Reviewable, reversible

Activity — what each run changed, and an Undo

Every agent run is logged with what it changed and why, and any committed change carries an Undo. Rejected proposals stay in the log too, so you can see what the agent wanted to do and did not. The library underneath is version-controlled, which is what makes undo cheap. The full trust model is on Sources & trust.

Try it yourself

Memento is in private beta. Join the beta, connect your own Claude or ChatGPT account, and the first capture is the same as the one in the clip above.