Cursor forgets your architecture every session. Here's how to stop it.

If you open Cursor, ask for a new feature, and immediately watch the AI suggest patterns you explicitly rejected last week, you are paying the Context Tax. It happens because chat-based AI assistants have no persistent memory of your project — only the current chat and whatever it picked up from the active file. On a senior salary that's ~15 minutes lost per session, ~62 hours per year, ~£4,650 of senior time. Here is the short answer, why it happens, and the persistence pattern that ends the ritual.

The short answer

Cursor only remembers what is in (1) the open chat, (2) the active file, and (3) any rule files you have committed to .cursor/rules/. It does not remember decisions, conventions, or yesterday's discussion. The fix is to externalise the decisions into directory-scoped .mdc rules and a single committed LEARNING_LOG.md — both of which Cursor loads automatically for the directories they apply to.

Why the chat alone can never fix this

Every Cursor chat is born empty. Even within a single session, the model has a finite context window; once you cross it, your earlier explanations start falling off the front. Open a new chat tomorrow and you start from zero again. Worse: the model has no idea which of your earlier statements were architectural decisions and which were exploratory thinking — so even if it had the transcript, it could not separate signal from noise.

"Just give it more context" is not a strategy. It is the bill for the Context Tax.

What persistence actually requires

Three things, in this order, all committed alongside your code:

1. Scoped rule files in .cursor/rules/ that activate only on the file types they apply to. A dotnet-di.mdc rule that loads on Program.cs and constructor edits — not on every prompt.
2. A Learning Log — a single LEARNING_LOG.md the AI is instructed to read at session start and append to whenever a non-obvious decision is made ("we use Result<T> not throw", "DbContext stays in Infrastructure", "Scrutor decorators preferred over manual factories").
3. A boundary-guardian rule that refuses cross-layer suggestions before they reach you — so you stop having to catch them in PR review.

That triple gives Cursor the closest thing to long-term memory you can build today, without waiting for any vendor feature.

The fix in 60 seconds

Drop one rule file in .cursor/rules/. The free arch-core-lite.mdc (linked below) is the boundary guardian — it activates whenever Cursor edits anything in your project and refuses cross-layer violations. The morning re-explanation ritual stops the same day you commit it. The kit's other rules layer on top: DI auditor, hallucination circuit-breaker, the Learning Log engine.

Want the full picture?

The long-form essay walks through the maths of the Context Tax, the four-rule architecture, and why this is a persistence problem, not a Cursor problem.

→ Read: The Context Tax — why every Cursor session costs you 15 minutes

---