The day I learned my assistant read 17% of its memory
My assistant keeps one memory file: lessons, decisions, what it has learned about me and the business.
When we audited the system in early September, that file was about 36,000 characters long. The chat was set up to load the first 6,000.
So 83% of what it “knew” never reached it. Nothing crashed. There was no error. It just answered with less, and nobody noticed until the audit.
The fix was one change: load the newest 9,000 characters instead. The lesson took longer.
A model only knows what is in front of it right now. The free starter shows you what to put there.
Is a context window the same as memory?
No. The window is what the model can see in this moment.
Memory is whatever system decides what goes back into the window next time: saved memories, project instructions, files.
That’s why a long chat can lose its beginning, and why a brand-new chat knows nothing about you unless something brings that information back in.
What happens when the context window fills up?
Something gets cut or squeezed. Depending on the app, older messages are summarized or dropped, or you’re asked to start a new chat.
You usually aren’t told what was lost. It’s one reason long sessions drift.
Does a bigger context window mean better answers?
Not on its own. A bigger window holds more, and more isn’t better.
An old version of your offer, repeated instructions and things that don’t matter all compete with the one thing you care about.
Every token also costs time and money.
My nightly review runs on an expensive model, so it has a rule.
Read the one summary my collector scripts prepared. Never re-fetch what they already pulled.
Images count too. In one photo review, a contact sheet of 30 thumbnails cost about 1,500 tokens.
When it matters
It matters for long documents, long chats, agents that run for many steps, and anything that loads files automatically.
It barely matters for a quick question in a fresh chat. If answers turn vague late in a long conversation, start a new one with a short summary of what’s been decided.
Choosing what goes into the window is its own practice, called context engineering.
For the difference between what you ask and what the model sees, read prompt vs context.
AI Foundations starts here: give AI a clear picture of your business in a form you can reuse, so you stop explaining from zero.
Background: Wikipedia: Context window