AI slang

Context rot.

What is context rot?

Context rot is the way an AI model's answers get less reliable as a chat or prompt context grows, even before the context window is full.

The slang landed because it names a familiar AI failure: the first answer is sharp, then the same chat slowly gets forgetful, fussy, or overconfident.

01 / Meaning

What the phrase means

Context rot describes decay inside the working material an AI model sees: chat history, pasted files, tool logs, instructions, memories, and retrieved passages.

The phrase is slang, but the behavior is measured. Chroma's 2025 report tested 18 models and found that performance became less reliable as input length grew, even on controlled tasks. Earlier long context research found a related pattern: models use information near the start and end of a long prompt better than material buried in the middle.

02 / How it shows up

How context rot appears and compounds

It usually looks like the model forgot something, but the quieter problem is attention getting spent on old clutter.

Buried constraints

A rule from ten turns ago still sits in the chat, but the model acts as if the latest error log matters more. This is why restating the current goal often works better than pasting another file.

Distractor buildup

Failed attempts, obsolete plans, console noise, and half-fixed bugs become part of the next prompt. The model starts optimizing against the sediment instead of the task.

03 / Examples

Concrete examples

In a coding agent session, context rot is the agent promising it removed an error, then repeatedly editing nearby code while the original failing condition stays untouched.

In a research chat, it is the model mixing your first rejected angle back into the final answer because that old instruction remains in the transcript.

In a document QA workflow, it is the answer missing a clause in the middle of a long contract while quoting nearby text with confidence.

04 / Distinctions

What context rot is not

Not simple truncation

Truncation means the input no longer fits. Context rot can happen while the prompt is still inside the advertised limit.

Not memory loss in the human sense

The model is not literally aging or forgetting. The joke borrows a human word for a machine problem: too much competing context makes the useful parts harder to use.

Not fixed by a bigger number

A larger context window lets you fit more text. It does not prove the model will reason over every part of that text equally.

05 / Practical guidance

How to work around it

Treat context like a workbench, not a warehouse. Keep only what the next decision needs.

Start fresh when the thread goes mushy

Carry forward the current goal, constraints, and verified facts. Leave failed plans and stale logs behind.

Retrieve narrow passages

For docs or contracts, load the paragraphs that answer the question instead of dumping the archive into the prompt.

Put live constraints at the edge

Repeat the rule that must not be missed near the latest request. Buried instructions are cheap to ignore by accident.

Related field notes

All field notes