Full writeup: the 4 things people call "memory", why RAG isn't one of them, hybrid retrieval, and the 5 things that break with real users.
Including the one that cost us real time: two writers, one store, or you get a split brain.
https://t.co/BX4Zmx5Q9C
Your AI assistant didn't forget your preferences. It never knew them.
A model API call is stateless. What feels like memory inside a chat is the client resending the entire conversation every single turn.
Turn 40 is the model reading turns 1-39 again. From scratch. For the ..
The fix is to stop asking the model to decide.
Read memory before the model is called at all. Put it in the prompt.
You pay latency on every turn — including the ones that didn't need it. In exchange there is no branch where memory gets silently skipped.