Fixing the Goldfish Brain Behind the Chat
The last biggest pain? The chat had the memory of a goldfish. Every message was treated like a brand-new convo. I couldn't remember what you said just seconds before.
The frontend was making up random session IDs, and the backend didn't get the memo; it created different ones. So, all the context was gone. I had to stitch that back together carefully without breaking the whole thing.
I also ran into several TypeScript mismatches between the UI and the database. It's nothing dramatic, just one of those "this should work… why doesn't it?" rabbit holes.
But it was worth it.
Now, chat sessions persist. The AI remembers up to 20 messages back. The session ID stays the same throughout, so conversations feel like conversations.
I also made a few quality-of-life tweaks:
- The chat is now on the homepage—no more landing on a boring dashboard.
- Messages render beautifully in Markdown (headings, bold, code, all of it).
- I cleaned up the schema on the backend, built solid API endpoints, and added fallback error handling for when things get weird.
What's next?
I'm rebuilding the RSS pipeline to pull in new episodes automatically. I don't think I need to keep the audio files. The approach will still be to scan the feed, grab new audio, and send it to Deepgram for transcription.
Then, it'll update the database and notify me when it's ready to chat.
All without me lifting a finger.
Thank you for reading. I appreciate your interest in this work. It has been an enjoyable challenge, and I’m still not finished. Unsure if I ever will be.
Talk soon,
Benoît