My personal reflections about web development, life, and more.
This Stopped Being About Data Engineering
The memory series ends where it started to become something else — behavioral data, preferences, intent, and the limits worth sitting with before building further.
You Don't Personalize the Model. You Personalize the Context.
One model per user sounds like the natural way to get real personalization. At scale, it's rarely the right question to be asking.
The Architecture That Actually Worked
Once conversations became state instead of history, the fix wasn't a better model — it was assembling context on purpose, layer by layer.
From History to State
Conversations aren't a list of messages to retrieve. They're a system that evolves over time — and that reframing changes everything.
RAG Sounded Like the Answer. It Wasn't the Whole One.
Retrieval looked like the fix for context overload. It was — until conversations turned out to be more than fragments.