Aaron Helwig

Software engineer building AI products
Work

Enlo

Personal assistant, in invitation-only alpha

Contents
  1. A relationship, not a pile of chats
  2. Talking to it while it works
  3. Done means delivered
  4. Memory that looks after itself
  5. How it fits together
  6. Where it stands

Enlo is the personal assistant I’m building: one assistant per person, with one conversation that carries on across Messages, the web and voice calls. I design and build all of it, from how a reply should read in iMessage to the runtime that keeps work alive underneath.

A Friday table for four, asked in Messages, with Nina’s walnut allergy already remembered.

A relationship, not a pile of chats

There is no “New chat” button and no model picker. You get one named assistant and one continuous conversation. Ask in Messages, keep talking on a call, open the web app later: the history, what it knows about you and the work in progress are the same everywhere.

  1. You ask in Messages for a table on Friday. Enlo already knows about Nina’s allergy.
  2. “Actually, 8:30.” The booking that is already running changes; nothing starts over.
  3. You open a browser voice call to discuss an option. Same assistant, same booking.
  4. Mid-call, you send the menu as a photo in Messages.
  5. The answer comes on the call, using the photo. That Sam is vegetarian now is remembered.
  6. You hang up. The booking finishes and arrives as a message.
  7. Days later, the same thread opens on the web with everything intact.
The evening from the phone above as one thread: the channel keeps changing, while the booking and memory underneath never reset.

That sounds like syncing, but most of it is judgment about where an answer belongs. If you send a photo to Messages while you are on a call, the useful reply is spoken, and the thread only needs a 👍. Enlo keeps where and when every message happened, so it can answer in the right place without repeating itself.

One history, not one copy everywhere.

Talking to it while it works

Real requests take time, and people keep talking. A message that arrives mid-task is treated as steering, not as the next item in a queue. A correction changes the work that is already running; an unrelated question gets answered alongside it.

  • “Actually, 8:30 works better”: correction steers the running task. Booked 20:30.
  • “Is Lupa the courtyard one?”: unrelated answered alongside it. Booked 20:30.
  • “Actually, 8:30 works better”: a plain queue waits for the old task to finish. Booked 20:00, then a second booking for 20:30.
Three messages arrive mid-booking. Only the last one waits in line.

Larger jobs are split between helpers, each with only the access it was granted. When a correction makes their work obsolete, the helpers pause so the main assistant can decide what still applies, and an outdated result is never delivered as current. Hanging up a call works the same way: the conversation ends, the work does not.

Done means delivered

“The model wrote a reply”, “Enlo published it”, “Messages accepted it” and “you saw it” are four different facts. Blurring them is how assistants end up claiming they told you something they never delivered, or sending the same thing twice after a hiccup.

Your phone is in a drawer

  1. Written: doneThe model produced a reply
  2. Published: doneEnlo committed it to the thread
  3. Accepted: doneMessages took it for delivery
  4. Seen: waitingOnly with evidence, like a read receipt

Accepted is not seen. Enlo does not tell itself, or you, that you already know.

One reply, four separate facts: written, published, accepted and seen.

The same care applies to work that changes things. In one recovery test, I stopped the backend and the controller in the middle of a real sandboxed program. When they came back, the original operation and its files were intact, and it had still run exactly once.

Ran once. Still ran once.

Memory that looks after itself

You should never have to tell an assistant to remember something. Enlo keeps the original history untouched and curates memory from it in the background, within a fixed daily budget per account, so maintenance can never crowd out the work you actually asked for. The memory notes in the phone above are the visible tip of that: Nina’s allergy recalled from months ago, Sam’s new diet picked up from a call.

How it fits together

The central decision was to write the agent loop myself rather than adopt a framework. The unusual parts of Enlo (one continuous history, steering, selective delivery, voice calls with their own lifetimes, recovery after failures) are exactly what a framework would decide for me. Models, app connections and sandboxes plug in as capabilities, so they can change without the conversation noticing.

  • Asked on a call, checked the calendar, answered in Messages
  • A spreadsheet in Messages, cleaned up in a sandbox, sent back
  • A question on the web, researched, explained on the next call
  • A quick question, answered straight away

Underneath, one history, memory and task list, stored durably.

Messages, voice calls and the web app all lead to one assistant, which reaches for models, apps, a code sandbox and research, with one durable history underneath.

Where it stands

The foundation works and is tested; breadth is what comes next. The longer aim is an assistant that reaches your own devices and takes on more of everyday life without you thinking about the machinery, while you stay in control of what it can touch.

Working in the alpha
  • Messages, web chat and voice calls, one history
  • Corrections that steer work already running
  • Helpers with scoped permissions, in parallel
  • Files in, programs in a sandbox, files out
  • Connecting apps from inside the conversation
  • Background memory, on a daily budget
  • Invitations, onboarding and billing
Next
  • Calling Enlo on an ordinary phone number
  • Slack, alongside Messages
  • Artifacts: briefs and dashboards with their own links
  • Working on your own computers and devices
  • Longer-running work environments
What works in the alpha today, and what comes next.

Earlier work deploying a personal AI for a private client shaped many of these priorities.

Curious about the alpha? Email me, though I can’t promise an invite.

Work