Solve "Remember to order milk" issue

[smian1]

This simple phrase is very powerful and important. This should never be discarded. How do we solve for this use case?

[maxfahl]

I’ve set up Zapier to use AI to analyze my speech in order to create and send to do items to my to-do app. I would love to be able to just speak something out loud and have the rest taken care of.

[e3e6]

Originally this device was intended to record dialogs, that's why it it's discarding short stuff like commands. The v2 has a button and a speaker, so it might work better for this type of flow, press a button, tell command, listen for response.

[andresgomezsar]

Originally this device was intended to record dialogs, that's why it it's discarding short stuff like commands.

The v2 has a button and a speaker, so it might work better for this type of flow, press a button, tell command, listen for response.

I think this is the way to Go, because the user experience is just better. Having the action button be the trigger for the short reminders gives users immediate feedback on whether their command was saved or their reminder was saved. That is a much better experience than just saying a keyword out loud and hoping that your system caught the keyword.

[maxfahl]

I'm quite new to the device. Is there something I need to do in order to make this work?

[beastoin]

Does the memory's Action Items / Events work for you @maxfahl

[maxfahl]

Im gonna create a webhook endpoint for raw audio data and transcribe it myself, since the built-in modems suck for Swedish. Then I will use OpenAI API to let the model decide on actions to call. One of them will be executing Apple Script to add to-do items to Things 3.

That's the plan anyway

[smian1]

Still an issue. Need better prompting to handle tasks. This should not be discarded.

[smian1]

@beastoin

[skywinder]

Hey 👋

According to the current prompt logic, the memory gets discarded based on this condition:
🔗 [Prompt in code]

 You will be given a conversation transcript, and your task is to determine if the conversation is worth storing as a memory or not.
    It is not worth storing if there are no interesting topics, facts, or information, in that case, output discard = True.

🧠 Logic:

• If ≤ 100 words → ask AI to evaluate content quality
• If AI says no interesting content → ❌ discard
• If AI says there is interesting content → ✅ keep

---

That said, the prompt itself could definitely be improved — someone could iterate on it via the linked logic

Personally, I think we should move these hardcoded prompts into a config/settings file, so users/devs can tweak them as needed. Otherwise, we’ll keep running into corner cases — and without tests, we won’t catch them all.

[skywinder]

✅ Suggested next steps:

• Move all prompts to a separate file
• Later: add tools for performance testing or let users adjust prompts directly
• This also partially addresses: [#1825](https://github.com/BasedHardware/omi/issues/1825)

Let me know what you think and which direction we should take! 💡