I and my friend work on similar project. lets connect!

[Awannaphasch2016]

My friend is working on pen.el project. https://github.com/semiosis/pen.el

we have discord here. https://discord.gg/g4Wn2gAJ

lets have a chat and trade our ideas. may be we can work together!.

[meedstrom]

I log in to Discord about once a year, but I'll drop in to brainstorm with you when I better understand your project and what it can do. Thanks for reaching out!

With Eva, before I resume coding, I want to go "back to the drawing board", because I feel that some insight is missing. A useful perspective came from asking myself what a virtual assistant should do for an user with BIMFH (ADHD, which I may or may not have, but it looks likely to help everyone):

• Do "beecatching": quickly and frictionlessly absorb information about the user's current thoughts and distractions, throughout the day. Maybe by acting as a "confessor" via prompts (where Pen or other LLM could help), and/or by inferring from computer activity. I'm thinking this information would be mapped to some loose concept of projects, of which Eva could generate new ones automatically depending on wording.
• Expire stale projects to save on the user's spoons (the mental energy needed to admit that the projects are stale and clean them up)
• Co-operatively start the user on projects -- instead of simply scheduling reminders, which to a BIMFH sufferer is like scheduling putting their hand on a frying pan. Reminders can make this sort of person just drop the system altogether. Instead, maybe Eva would do a little interaction where the user picks from a small list of options or supplies their own option, then follows their activity for a while to check that they've really begun working. This is another place Pen could help. When the user does not work on the stated project, Eva would react without judgment, no "Back to work!" attitude but more of a therapist's "what are you thinking about?". Aside from the immediate benefit of making the user self-aware in the moment, prompting braindumps from the user at crucial moments like these could prove very useful later when they talk to a real therapist.

[xeruf]

I currently run https://activitywatch.net for tracking all computer activities, but never get around to reviewing its stats - maybe something sensible to integrate?
I actively use Stretchly and a self-written script called netkeeper to limit my internet access and nudge me to use my time on the computer intentionally, though I am not completely satisfied with this combination yet.

Oh, and AI has become very useful within the recent years, maybe it is time to make the assistant actually intelligent :)

[meedstrom]

Oh, and AI has become very useful within the recent years, maybe it is time to make the assistant actually intelligent :)

Even more true now in 2026 :) Did you do anything more with ActivityWatch? @xeruf

For starters, it seems relatively simple to set up a MCP to give a LLM access to all the text and data you've ever produced, as that'll surely help it say more useful things to you. But that's still just a LLM, not a full assistant.

Zooming out, it seems interesting to ask what parts of your assistant should be driven by:

• direct interactions with a LLM, like a chat or popup
• the LLM "talking to itself" to reason out what to do
• classic software
  • at minimum, some script needs to regularly poke the LLM into reviewing the current situation to decide whether there is some action it should take, right? A good assistant won't wait for the user to talk to it.
  • of course, ActivityWatch and similar for data
  • the UI/UX of how you'll interact in practice, it's a bit of software to implement, but it's also a huge design question. Do you "complete missions" for the assistant? What happens if you neglect or repeatedly decline a mission? See my concepts of "user excursions" and "loosely scheduled actions"
• actual mathematical models, like the Bayesian time-series stuff in Eva's README, to predict some things
  • the utility may not be obvious since we're all so enamored with what LLM and "AI" can do right now, but strictly speaking, a pure LLM has no idea how to assist you if it has no data about what's going on today (and it should not rely 100% what the user says, let alone assume it has access to the user's attention at the time that it would matter), and a predictive model is in a sense just a way to conjure additional data, some of which would be very valuable in decision-making