06 of 12
Memory
Scratch work on a napkin is for this problem. A notebook is for the next one. If you throw the napkin away, tomorrow you start from zero. Memory is the notes you keep on purpose for later.
Early chat systems lost the thread once the box filled and greeted you like a new person. Teams first crammed more text into the question. That helped for a file you had open. It did not help for a decision made last month. What coding assistants actually shipped is files the model can read and write across sessions: a note about the project, a list of commands that already broke, a short brief a person left for it.
Why not just a bigger window?
Google and others advertised windows of a million words. The demo is a fact planted in a long file, which the model can find. A real repo is four hundred files and one constraint from last quarter. Cursor still makes you keep a notes file because stuffing the repo into every question is slow, expensive, and still misses the constraint.
Is fetching a wiki page memory?
Glean and Notion will pull a few wiki pages into the question before answering. That is a guess about which page matters. If two pages disagree, or the decision lived in a meeting, the answer can still be wrong.
Why does it still get last month wrong?
ChatGPT Memory, launched in 2024, writes a profile from the chat. It keeps a plan you abandoned unless you delete it. A notes file in Cursor has the same rot the day the API changes.