Ask people what they want from an AI agent in their messages, and the answer is often not "more features". It's "remind me". Builders of text-based assistants have said publicly that simple reminder requests are among the most common things people ask for. People in ADHD communities describe the same instinct from the other direction: they text themselves things, and pin those chats, so they don't forget.
Reminders sound like the easy part. In practice, they're where AI assistants lose people's trust fastest.
Three ways reminders fail
1. They silently don't fire. In public forums, users of AI assistants describe scheduled reminders and automations that sometimes just don't trigger, and that get confused when time zones change. A reminder that doesn't arrive is worse than no reminder, because you stopped holding the thought yourself.
2. They're confidently wrong. Users also describe assistants saying a task was done when it wasn't, or setting something at the wrong time while insisting it was right. The bug isn't only the wrong time. It's that you weren't told.
3. They're one ping you can swipe away. Many people ask for reminders that keep going until something is actually done. Parents talk about wanting reminders that reach family members who ignore app notifications. A single notification is easy to dismiss and forget.
None of these failures is really about intelligence. They're about honesty and follow-through.
What we're building instead
Moralie is a personal AI agent we're building to work inside a Telegram chat. It isn't available yet, and everything below is what we're designing for, not something you can use today.
- A confirmation you can check. When you ask for a reminder, Moralie should repeat back exactly what it saved, including the date, time and time zone, so you can catch a mistake right away. Something like: "Got it: Tuesday, 3:00 PM your time. Reply 'change' to fix it."
- No claims of "done" without a record. We're designing Moralie to only say something is saved once it actually is. If it couldn't do something, it should say so plainly. "I couldn't set that" beats a confident wrong answer.
- Dates checked by ordinary code. Natural language goes in, but the date it turns into should be validated by plain, deterministic code before anything is saved, rather than trusting the language model's interpretation.
- Opt-in follow-ups. For people who want it, a reminder that checks back until you reply "done", with an easy way to stop.
- The chat you already check. No new app to open and no dashboard. The goal is for your first message to be the setup: no forms, no API key.
We'd rather ship a small set of things that behave honestly than a long feature list that occasionally lies to you.
If you're building reminders too
A few lessons that apply to any assistant, not just ours:
- Echo back the parsed time, with the time zone, every time.
- Never let the model's words be the source of truth for whether something was scheduled. Check the actual record.
- Treat a reminder that didn't fire as a serious bug, not an edge case.
Help us get it right
Moralie's soft beta is coming soon, and we plan to open it in batches. The first people on the waitlist will get invites first. We can't promise a spot or a date, but we'll be straight with you about where things stand.
👉 Join the waitlist: https://moralie.pages.dev?utm_source=devto
What's the reminder an assistant has let you down on? Tell us in the comments.
— The Moralie team
Top comments (0)