TL;DR
An AI's past answers in a long conversation aren't always an asset. Researchers from MIT and IBM examined real multi-turn conversations and found that if you strip every previous assistant answer out of the context, response quality either holds steady or... goes up. The culprit is "context pollution": the model latches onto its own earlier answers and drags their errors, hallucinations and style into every new reply.
Sound familiar? You spend several messages editing a draft with Claude, and the model stubbornly clings to the structure of its first version even though you've asked three times for a rewrite. Or you ask for code using a new library, and it still pulls arguments from the old implementation. That's not a bug — it's a systemic trait: the model sees its previous answer and "trusts" it, even when it was off.
It turns out 36% of questions in real conversations are standalone: they don't depend on anything the model said before. Another third are edits with specific instructions that can also be answered "from a blank slate." In practice, only about a third of conversations actually need the model to remember its own past words. For everything else, a fresh chat works just as well — and often better.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
This isn't a step-by-step prompt — it's a framework for managing conversation context. The main tool is classifying your next question correctly before you ask it.
STEP 1: Identify the type of your next question
↓
"New request" → Start a new chat (or tell the model to ignore the past)
"Edit with instructions" → You can continue, but copy over only what's needed
"Reference to a past answer" → Keep the full context
STEP 2: If you see the model "dragging along" past mistakes →
Say it explicitly: "Don't rely on your previous answer. Solve it from scratch."
Or: start a new chat and give only the problem statement
Every step happens in a regular chat — no code required.
Example in Practice
Task: You're editing an investor pitch deck with Claude. In the first draft the model went with an overly aggressive tone of voice — like it's selling vacuum cleaners, not SaaS. You asked for a rewrite, it softened a little, but the salesman vibe stuck. Now you want to write the team section — and you don't want the old tone leaking in.
Prompt:
Write the "Team" section for a pitch deck.
Context: we're building a B2B SaaS that automates procurement for retailers.
Round: pre-seed, raising $1.5M.
Team:
— Alex Morgan, CEO: 8 years in procurement at Kroger, built supply systems from scratch
— Emma Collins, CTO: former Google engineer, specializes in ML pipelines
— Daniel Hayes, CMO: grew Instacart's grocery vertical from 0 to 2 million orders
Tone: calm, confident. The facts speak for themselves — no need
to dress them up with adjectives. Write like a YC founder, not a salesperson.
Ignore all previous versions of the text from our conversation.
Start from a blank slate, relying only on the brief above.
Result: The model writes the team section in the right tone — without the "pushy salesperson" from earlier iterations. The explicit "ignore previous versions" instruction breaks the chain of context pollution. The section stands on its own and won't carry stylistic artifacts from early drafts.
Why It Works
The model doesn't "forget" its past answers — it trusts them. When a previous answer sits in the context, the model treats it as part of the "truth" about the task. If there was an error, it doesn't double-check — it builds the error into the next answer. That's how a mistake hops from turn to turn.
Long context has a price. The more text in the conversation history, the harder it is for the model to "zero in" on the right spot. This isn't a metaphor — research shows models literally do worse when the relevant fragment is buried in a huge context. The AI's own past answers are the fastest-growing source of noise in a conversation.
The fix is controlling what the model sees. Three tools: (1) a new chat for new topics, (2) an explicit "don't rely on your previous answer" instruction, (3) pasting only the piece you need instead of hauling the whole history along. The lever is simple: the less of its own past text the model sees, the lower the odds it brings a mistake back from there.
Control levers: - "Ignore your previous answer / start from scratch" — the most direct way to break the pollution - New chat — the nuclear option, always works, but you lose useful context - Paste only the needed fragment — surgical: give the model only what the current request requires - Explicit rephrasing — instead of "fix this," write "do X using the conditions from my first message" — the model leans on you, not on itself
Prompt Template
Diagnosing the Request Type
Look at my next question in this conversation and determine its type:
1. "New request" — a standalone task, unrelated to your previous answer
2. "Edit with instructions" — I'm changing something specific (different language/style/format)
3. "Reference to an answer" — I need part or the gist of a past answer
My next question: {question}
Answer in one word: the type + one line explaining why.
Resetting Context Pollution
Solve the task from a blank slate. Don't use previous versions from our conversation.
Task: {task}
Conditions: {key parameters — what must be preserved}
Format: {what you want to get}
What to fill in:
- {task} — a concrete wording, as if you were typing into a new chat
- {key parameters} — what genuinely needs to be remembered (context, constraints, data)
- {what you want to get} — the answer format
🚀 Quick start — paste into the chat:
Here are two templates for managing conversation context.
Adapt them to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask what the task is and which parts of the earlier conversation need to be kept — because that determines whether you should reset the context or just sharpen the instruction.
Limitations
⚠️ Doesn't work for iterative editing: If you're literally saying "here's your text, fix this part," you need the previous answer. Resetting the context here will break the workflow.
⚠️ Reference questions need history: "Explain the second point in more detail" — without the previous answer, the model has no idea what "the second point" is. Full context is mandatory here.
⚠️ The effect varies across models: Stronger models (GPT-5.2, Claude) handle polluted context slightly better than smaller ones. But context pollution shows up in all of them — top models included.
⚠️ "New chat" isn't always the best choice: If you need a chunk of a previous answer, it's easier to copy it into a new chat explicitly than to haul the whole history along. But if you don't need the past at all, a new chat is the safer bet.
How They Studied It
The MIT and IBM team took real conversations from two datasets — WildChat and ShareLM — focusing on technical chats (code, math). The idea was simple: run each conversation in two modes — with the full history (including the AI's past answers) and with only the user's turns (AI answers removed and replaced by a placeholder [Response provided]). Then compare quality.
GPT-5 served as the judge — it looked at both answer versions and picked the better one. For robustness they ran two judges: one saw the full history, the other only the user's turns. The results agreed: removing the AI's past answers doesn't hurt quality for most turns, and sometimes improves it. Most surprising of all, this held even for GPT-5.2 (one of the strongest models at the time of the study).
The team also trained a simple classifier to predict whether history is needed for a given turn. It preserved 95% of full-context quality while delivering 30% token savings. Key insight: the rule "drop history only on new requests" works considerably worse than judging each specific turn — because even follow-ups are often solvable without the AI's history.
Original from the Paper (optional)
Examples of context pollution from the study:
Conversation: t-SNE vs. UMAP Code
What happened: The model incorrectly carries over UMAP-specific arguments
from an earlier turn when asked to rewrite the code using t-SNE,
resulting in a bug.
Conversation: Book Recommendations
What happened: The model hallucinates book recommendations in an earlier
turn (real books, but not ones that satisfy the user's request)
and persists in mentioning them in later turns.
Conversation: Model Temperature
What happened: The model introduces a mathematical formula in an earlier
turn and then applies it incorrectly in a later turn.
Conversation: Reflecting on Python Code
What happened: Instead of following a new user instruction, the model
continues generating content in the same tutorial style as an earlier turn.
Context: The researchers looked for cases where the "assistant-omitted" version clearly beat the full context, then manually dissected those conversations to find the point where the pollution started.
Adaptations and Extrapolations
💡 Adaptation: Auditing the Current Conversation for Pollution
If the conversation is already long and the model seems "stuck," ask it to diagnose itself:
Look at our conversation. Are there assumptions, mistakes or
stylistic choices in your previous answers that you keep
carrying into new answers without good reason?
If so, name them explicitly. I'll tell you which ones to drop.
The model won't always find the problem, but it often "sees" patterns worth resetting on its own.
🔧 Technique: An Explicit "Checkpoint" at the Start of a New Step
Instead of continuing a long conversation, periodically reset with an explicit summary:
Let's start this step from a blank slate.
Here's what to keep from our conversation:
— {key fact 1}
— {key fact 2}
— {condition / constraint}
Ignore everything else from your previous answers.
Task for this step: {new task}
This works especially well in long writing sessions, strategy work and multi-stage analysis — where every 5-7 turns it pays to "zero out" and restate the context by hand.
Resources
Do LLMs Benefit From Their Own Words? Jenny Y. Huang, Leshem Choshen, Ramon Astudillo, Tamara Broderick, Jacob Andreas Massachusetts Institute of Technology, MIT-IBM Watson AI Lab, IBM Research Preprint, 2025
Related work mentioned in the paper: WildChat (Zhao et al., 2024), ShareLM (Don-Yehiya et al., 2025), ERGO (Khalid et al., 2025)