"When did I last buy 10kg of garbanzo beans, and from where? What price did I pay?"
And get an answer like:
"Actually, you didn’t buy 10kg; last time you only bought 5kg from X shop at Y price. Based on your past consumption, your stock will likely run out by next week. Should I set a reminder for today, after your gym session, to pick some up? (X shop is on your way home from the gym.)"
This level of contextual response would be incredibly useful. These days, I bulk order everything thanks to my streamlined note-taking and reminder setup. I’m surprised there isn’t already an open-source tool that works this well.
I keep my day organized with simple methods. During my morning walks, I plan out my tasks, priorities, and schedule—talking through everything in my head. These thoughts are then transcribed using speech-to-text and sent to an LLM. Since LLMs aren't great at remembering specific facts or handling complex relationships, I pair it with a knowledge graph to keep everything organized.
This setup generates reminders, creates schedules, and flags conflicts where I can reschedule or drop tasks. I dislike most conventional note-taking or reminder apps, so I stick to plain text files stored across Dropbox, a Raspberry Pi home server, and cloud storage like S3.
To keep me on track, I’ve built a custom notification system that sends reminders through text, email, Telegram, and WhatsApp. These notifications continue—staggered across platforms—until I acknowledge them. Since I rarely use my phone, I rely heavily on a smartwatch that receives SMS notifications. It’s a game changer: with its own SIM and long battery life, it costs me almost nothing—just $30 a year.
I avoid traditional apps for adding new information. Instead, I use a private Telegram group with a bot for input. Messaging in this group has become the easiest way for me to update my system, and I’ve grown to rely more and more on Telegram bots for this reason.
For example, yesterday the system reminded me to check my solar batteries. Months ago, I had told it that I watered them, and it automatically followed up at the right time. It’s these small, automated details that help me stay on top of long-term tasks.
I’m using Gemini Flash (a dirt-cheap, fast LLM), Neo4j, and Whisper, all tied together with Python glue scripts to make this work. Maybe someday I’ll have hardware powerful enough to run a local LLM, but for now, this setup is more than good enough.