Why Cursor 'forgets'
Cursor, like every AI coding tool, is stateless: each session starts without memory of the last. It only knows what's in its context window right now, and that window is finite — so prior decisions and conventions vanish unless something re-supplies them.
It's not a defect you can prompt around. The teams that get the most from Cursor design for it rather than fighting it.
Use project rules and a context file
Put your durable conventions and constraints in Cursor's project rules / a root context file. This is the equivalent of a CLAUDE.md: architecture overview, naming conventions, the libraries you've standardized on, and the things never to touch.
Now every session starts informed instead of blind, and you stop re-explaining the same ground rules.
Let retrieval do the heavy lifting
Don't paste your whole repository into the chat. It's slow, costly, and buries the few files that matter. Let Cursor index and retrieve the relevant code for the task at hand.
Point it at the specific files and let its codebase indexing pull the rest. Relevant context beats exhaustive context every time.
Capture decisions, not just code
The most expensive context to recover is a decision and its rationale. Weeks later nobody remembers why a pattern was chosen, so Cursor (and your teammates) reopen settled questions.
Record decisions where Cursor reads them — a short entry on what was decided and why — and the agent stays consistent with reality.
Keep it current
A stale rules file is worse than none: Cursor will confidently follow conventions you've abandoned. Review the context like code and update it as the project changes.
Automate the upkeep where you can so it never depends on someone remembering to edit a file.