Memento Mail, first called Needle, is earlier work and no longer in active development. This page describes what we built and what I learned from it.
Context and problem
The inbox is where work goes to disappear. A proposal is sent and never followed up. A reply is promised and never written. An intro thread stays open until it is too awkward to answer. A deadline sits inside an attachment nobody opens twice.
Email tools optimize for reading and writing messages. Almost none of them care about the state in between: the commitments, requests and decisions that stay unresolved after the message is read. That unresolved layer is where deals stall and trust leaks.
What the product did
Memento Mail was an AI-assisted email workspace. It read the inbox the way an operator would, scanning messages and attachments for the work hiding inside them: follow-ups that were never sent, replies that were promised, threads left open, deadlines and requests buried in documents.
What it found became a working layer above the inbox: a live list of unresolved work, each item tied back to the thread it came from. It was not an email summarizer. Summaries compress what happened; Memento Mail surfaced what still had to happen.
My role
Co-founder and product engineer. I drove the product, the AI workflows that detected unresolved work, and the positioning: how the product explained itself to the people whose inboxes it entered.
System architecture
- Email threads
- Attachments
- Ongoing conversations
- Scans for unresolved work
- Detects promises and requests
- Finds open loops
- Follow-ups that never went out
- Promised replies
- Open threads and deadlines
Mail flowed through an analysis layer that looked for unresolved states rather than topics. Everything it found landed in the working layer with its source attached, so acting on an item took you straight back to the thread where the work lived.
Engineering decisions
Unresolved work is the unit, not the summary
The obvious AI email product is a summarizer. We deliberately built around a harder unit: the unresolved state of a thread. Detecting that a reply was promised and never sent is much harder than compressing text, and much more valuable, because it points at an action instead of a recap.
Precision over volume
A layer that cries wolf gets closed by Friday. We tuned it to flag fewer, truer items rather than everything that could possibly be a task. The cost was that some real work slipped through early; the benefit was that people trusted what the layer showed.
Attachments are first-class
Commitments and deadlines often live inside documents, not in the message body. Attachments were treated as part of the conversation and analyzed for the same unresolved work.
Product decisions
A layer, not another inbox
The product never tried to replace the inbox or become another place to live. It sat above the mail and pointed back into it. That framing shaped everything from the interface to the way we talked about it.
From Needle to Memento Mail
The product started as Needle. As it moved from a detection tool toward a full email workspace, the name changed to Memento Mail.
What I took from it
Memento Mail taught me that trust is the real product when AI touches private data: people only let a system into their inbox if it is precise and explains itself. That question, how to use AI without giving up control of your data, is the one I now work on full time at Derbent.