Rex: My Memory Assistant
A Telegram bot that read my Obsidian notes, wrote questions from them and asked me those questions on a spaced repetition schedule. It worked. I stopped using it for reasons that had nothing to do with the code.
Can the job of remembering what you have read be handed to a machine?
The idea was unlimited memory. I read and take notes much faster than I revise them, so notes pile up in Obsidian and quietly stop being worth anything. This was meant to close that gap. It read the vault, worked out what I was likely to forget and when, and asked me about it on Telegram before I lost it.
How it worked
A parser walked the Obsidian vault and pulled out the material that could be turned into questions. A generation step wrote the questions. An evaluator marked my answers. A scheduler decided when each note came back, widening the interval each time I got something right, following the Ebbinghaus forgetting curve. A second scheduler handled the messages the bot sent without being asked for them.
Note and schedule records live in SQLite. Runtime state, meaning pending reminders, scheduler state and conversation history, sits in JSON files on disk alongside it.
Why Telegram
The API was the easiest to set up, which is the whole reason. I tried WhatsApp first and the setup threw up bug after bug, so I stopped fighting it and moved to the platform that let me get to the actual problem faster.
Why I stopped
Two reasons, and the code was neither of them.
The first was question quality. Turning a note into a question worth answering is much harder than it looks. The model I could afford to run was not good enough at it, and a weak question does more damage than no question, because answering it teaches you nothing and you stop trusting the thing that asked.
The second was money. Better models existed and would have written better questions. This is a background job that runs constantly over a pile of notes that only grows, so the bill scales with how much I read, and I read a lot. On a school student’s budget the arithmetic did not work.
What I would keep
The problem is real and I still have it. My notes still outrun my revision. What I got wrong was assuming the hard part was the scheduling, when the hard part was writing a question worth being asked.
The colophon explains how I build software, and where AI was involved.
- Software
- Spaced repetition
- Python