What Is Context Rot?
Context rot is the measurable decline in AI output quality as a conversation or input grows longer. The model starts forgetting instructions, drifting off topic, and contradicting things it said earlier. It is not a bug in your prompting. It is a documented property of how large language models handle long inputs.
If your AI outputs feel inconsistent, sharp at the start of a session and sloppy an hour in, you are not doing anything wrong. You are watching context rot happen in real time.
Research by Chroma published in 2025 tested 18 frontier models, including GPT and Claude families, and found that every single one degraded as input length increased, even on tasks as simple as retrieving a fact from earlier in the input. The bigger context windows get, the more this quietly matters.
Why Does AI Get Worse as a Conversation Grows?
Models pay uneven attention across their context window. Content at the beginning and end of the input gets weighted heavily, while material in the middle gets under-weighted. Add in accumulated clutter from a long back-and-forth, and the model's effective grasp of your conversation shrinks even though everything is technically still "in" the window.
The academic version of this is the "lost in the middle" effect, documented by Stanford researchers in a widely cited study: models retrieve information far more reliably when it sits at the start or end of the input than when it is buried in the middle.
There is also a noise problem. A two-hour working session accumulates dead ends, rejected drafts, and abandoned tangents. All of it stays in context, and all of it competes for the model's attention against the things you actually care about.
This is why "just paste everything in, the window is huge now" is only half true. Capacity and attention are different resources. The window grew. Attention did not grow with it.
How Do You Know Context Rot Is Setting In?
The reliable warning signs: the model re-asks questions you already answered, ignores a formatting rule it followed earlier, blends details from two different topics discussed in the same chat, or produces answers that get vaguer while getting longer. Any one of these means the conversation has passed its useful length.
A concrete example: you set a style rule in message one ("keep all copy under 50 words, no exclamation marks"). For the first ten outputs it complies. By output twenty, exclamation marks are back. The rule did not disappear from the window. It disappeared from attention.
Practitioners often respond by scolding the model ("I told you no exclamation marks!"), which adds more middle-of-context clutter and makes the underlying problem slightly worse. The fix is structural, not motivational.
How Do You Fix Context Rot?
The fix is active context management: deciding what stays in the conversation, what gets summarised, and when to start fresh. Five habits cover most situations, and none of them require technical skills.
1. Treat sessions as disposable. One task, one conversation. When the task changes, start a new chat rather than steering an old one in a new direction.
2. Summarise and restart. When a long session starts drifting, ask the model to compress the session state into a handover note, then paste that note into a fresh conversation. You keep the decisions and lose the clutter. The exact prompt is in the next section.
3. Front-load and end-load what matters. Put standing rules and key reference material at the very start. When asking the final question in a long input, restate the critical constraint at the end. The middle is where instructions go to be ignored.
4. Keep a standing brief outside the chat. Store your style rules, project background, and preferences in a document, or in a feature built for this such as Claude's Projects or ChatGPT's custom instructions. Reference material that lives outside the conversation cannot rot inside it.
5. Prune before you paste. Feeding the model 40 pages when 4 are relevant does not make answers better. In long-context testing, more distractor material consistently lowered retrieval accuracy. Relevance beats volume.
The Summarise-and-Restart Prompt
This is the single highest-value habit on the list. Run it whenever a productive session starts feeling muddy, typically after 20 to 30 exchanges.
Try This Prompt:
"This conversation is getting long and I want to continue in a fresh session. Write a handover briefing for the next session that includes: (1) the goal we are working toward, in one sentence; (2) every decision we have made so far, as bullet points; (3) all standing rules and preferences I have given you, stated exactly; (4) the current state of the work, including the latest full version of any draft; (5) the immediate next step. Do not summarise loosely. Anything ambiguous, quote my original wording. Output only the briefing, ready to paste."
Then open a new conversation, paste the briefing as your first message, and add one line: "Confirm you have absorbed this briefing, then continue with the next step." The confirmation forces the model to actually process the handover before working.
Item 3 matters most. Rules restated at the top of a fresh session sit in the highest-attention position, which is exactly where degraded rules from the old session were not.
What Are the Common Mistakes When Managing Long Sessions?
Three traps catch even experienced users. Each one feels productive while quietly making outputs worse.
Marathon chats as a badge of honour. A 300-message conversation is not a rich collaboration. Past a certain length it is a liability. The model does not get to know you better over a long chat; it gets a noisier window.
Restating instead of restarting. Repeating your rules every few messages does help attention briefly, but it also duplicates content and bloats the window faster. One clean restart beats ten reminders.
Trusting recall over verification. In a long session, never assume the model still remembers a constraint from an hour ago. Before a final deliverable, ask: "List the rules you are currently following for this task." If the list is wrong, restart before polishing.
The Bottom Line
Context rot is not a flaw you can prompt your way around with cleverer wording. It is a structural property of long inputs, and the practitioners who get consistent results are the ones who manage the window deliberately: short sessions, clean handovers, rules kept at the edges where attention lives.
Master the summarise-and-restart habit this week and you will feel the difference in a day. Technology should adapt to how you work, not the other way around. UD has walked beside Hong Kong businesses for 28 years, making technology feel human.
How Deep Does Your AI Knowledge Actually Go?
Context management is one of the skills that separates casual AI users from power users. Take the AI IQ Test to find out where you stand, and when you are ready to build reliable AI workflows into your daily operations, we'll walk you through every step, from setup to deployment.