Context windows aren’t memory

A bigger context window is not the same as a system that remembers. It’s a bigger desk, not a filing cabinet — everything you put on it is in view at once, and everything you don’t is simply gone.

Retrieval beats stuffing

The instinct is to paste in more: the whole document, the whole history, the whole knowledge base. It works in a demo and falls over in production, where cost, latency, and relevance all degrade as the prompt grows. We retrieve only the few things that matter for this turn and leave the rest out.

Forgetting is a feature

A useful agent decides what to keep. Summaries, not transcripts. State, not logs. The skill isn’t remembering everything — it’s knowing what’s safe to drop so the model spends its attention where it actually counts.

Memory is a system, not a setting

Persistent memory lives outside the model: a store you write to, retrieve from, and can inspect. The context window is just the working surface where the right pieces meet. Treat them as separate and the whole thing gets more reliable — and a lot cheaper to run.