Long context windows still rot

Modern models advertise million-token windows. You can read all of Hamlet and have room left over. So why does AI start making weird decisions four hours into a long Cursor session — using a deprecated helper you removed two hours ago, mixing up User and UserProfile, introducing a bug you already fixed?

Run the editor on the right. That session log is a toy version of what Cursor or Claude Code is carrying around — a growing pile of turns, each one full of tokens, each turn pulling the model's attention in a slightly different direction.

The window doesn't truncate — that part works. But the model's attention does dilute. By turn 6, the model has to pull signal out of a soup that includes a deleted helper from turn 1, a stack trace from turn 3, and three different refactor attempts from turn 5. Even a model with a million-token window has to choose what to weight, and choices get noisy when the inputs do.

The two failure modes

Long sessions decay in two distinct ways. Knowing which one you're in tells you what to do.

Mode 1: stale context. The model is using something that's no longer true. You renamed Order to OrderRecord an hour ago. The model still calls it Order. The fix isn't a better prompt — it's a fresh session with the current file pasted in.

Mode 2: drift. The model started solving the right problem and slowly slid sideways. You asked for a bug fix, it fixed the bug and also rewrote three unrelated functions. The fix here is the plan-first prefix (next paragraph) before the next prompt.

The plan-first prefix

This is the single highest-ROI prompt move you'll learn this lesson:

Before writing any code, write a 5-line plan for what you'll change
and which files. Wait for me to approve. Then implement.

Drop that into any non-trivial request and three things happen:

1. The model articulates its intent in a form you can read in 20 seconds, before any code gets written.
2. You catch wrong assumptions cheaply (you say "no, don't touch X" and it adapts before generating).
3. When the diff lands, you can compare it against the plan — drift becomes visible.

Use the prefix any time the change touches more than one file or involves logic you can't fit on one screen.

When to start fresh

Three signals you should kill the session and open a new one:

1. The model references a file or symbol that no longer exists. Pure stale context. New session, paste the current file, restate the goal.
2. It's hallucinating function signatures. Especially common with library upgrades. The model is averaging across versions in its training data. New session with the current requirements.txt or package.json pasted in.
3. You've said "no, that's wrong" twice in a row. The session has a wrong working theory baked in. Anything you say now fights that theory. Start over.

What about Claude Code's /clear and /compact?

Same idea, lower friction. /clear starts a fresh conversation with empty context — your CLAUDE.md, skills, and settings.json persist because they live on disk, not in the chat. Use it freely.

For mid-task context relief without losing the thread, reach for /compact instead — it summarizes the conversation so far and keeps going. /compact is the more idiomatic move when you're not done but the buffer is heavy. /clear is for "I'm starting something new."

You're not losing progress either way — the files are the progress. The chat is a means.

How AI specifically gets this wrong

When AI suggests "let me continue where we left off" after a long gap, that's a tell. The model has no memory between sessions; it only sees what's in this turn. If "where we left off" isn't visible in the current conversation or the current files, it's making something up. Don't reward that. Restate the goal as if from scratch.