When your AI chat gets lost, don't patch it — restart it clean

Overview

In a study of 15 models across more than 200,000 simulated conversations, the same task drip-fed over several turns scored 39% worse, on average, than when the whole task was handed over in one complete message. The size of the drop isn’t even the interesting part. What broke is.

The models didn’t get less capable — they got less reliable. Aptitude fell about 16%; unreliability more than doubled, a 112% rise. And once a model took a wrong turn early, in the researchers’ words, “when LLMs take a wrong turn in a conversation, they get lost and do not recover.”

So when a long chat starts going sideways, your instinct — fire back another correction in the same thread — is the wrong one. What you’ll do differently: stop, consolidate, and restart clean.

The content

The study behind this (“LLMs Get Lost in Multi-Turn Conversation,” May 2025) is older than most of what this site cites, and that is not a weakness here: it keeps being re-confirmed. Through mid-2026, separate teams have found the same decay on newer models — as user intent evolves across turns, and in real multi-turn coding sessions. The 2025 study stays the anchor because it’s the one that measured the fix, not just the failure. It ran each task two ways. Sharded: revealed piece by piece across turns, the way a real conversation actually goes. Concat: every one of those pieces handed over in a single consolidated turn. Sharded is where the 39% drop lived. Concat recovered almost all of it — “Concat performance averaging 95.1% of the Full counterpart.” It wasn’t the amount of information that hurt the model. It was the drip.

That points to the fix, but there’s a tempting shortcut to rule out first: when a chat derails, why not just restate everything in your next reply, same thread, one clean summary? The researchers tested that — a recap turn, and repeating your points as you go. Both help a little: “improvements over Sharded,” but “the performance still lags behind Full or Concat,” and repeating your turns recovers only 15–20% of what was lost. The model’s earlier wrong turns are still sitting in the thread, still anchoring it.

Hence the move. Consolidate everything you now know — the full goal, every constraint, every correction you’ve had to make — into one clean message, and paste it into a fresh conversation. You’re recreating the Concat condition by hand: all the information, none of the drip.

One honest boundary, because it changes how you apply this. The evidence is comparative, not magic. What’s proven: one consolidated turn beats restating in-thread, which beats carrying on. What’s not proven is that a new thread is strictly required, or that any restart rescues a bad prompt. The lever is the quality of the consolidation; the fresh thread just clears the model’s earlier wrong turns so the clean version can land. And one asterisk on the numbers: while the problem keeps being re-confirmed, the remedy’s size hasn’t been re-measured — the near-total recovery from consolidating is a May 2025 measurement on that generation of models. Treat the effect as standing, and the exact recovery figure as a dated snapshot.

Try it

The next time an assistant — ChatGPT, Claude, Copilot, Gemini, an internal wrapper — has clearly drifted, and you’re two or three corrections deep and still off, don’t send correction number four.

1. Stop replying in the derailed thread.
2. Write ONE fresh message, from scratch, that includes:
   - the actual goal (what "done" looks like)
   - every constraint you've discovered since you started
   - every correction you've had to make along the way
   written as if the assistant had never seen the conversation.
3. Open a new chat and paste that single message in.

Where it breaks: if your consolidated message is as vague as your first prompt was, a fresh thread won’t save it — the restart helps because it drops the model’s earlier wrong assumptions, not because “new” is magic. The work is in the consolidation. And if a long chat is still behaving, leave it — this is a recovery move, not a routine to run on every conversation.

Additional reading

• “LLMs Get Lost in Multi-Turn Conversation” (Laban et al., Salesforce AI Research & Microsoft Research, 9 May 2025) — the driver: 15 models, 200,000+ simulated conversations, a 39% average drop across six generation tasks that splits into roughly 16% lost aptitude and a 112% rise in unreliability; “when LLMs take a wrong turn in a conversation, they get lost and do not recover.” Concat recovers to 95.1% of full; in-thread repetition mitigates only 15–20%.
• “LLMs Get Lost in Evolving User Intent” (Tack, Laban & Neville, 22 July 2026) — the recency check on the problem: strong single-turn performance “does not transfer to the evolving-intent setting, with substantial drops across model families.” The models behind that sentence are the current ones — GPT 5.1 through 5.5, Gemini 3.1 Pro, Grok 4.20, Kimi K2.5 and K2.6, Mistral Large 3 and DeepSeek V3.2 (the per-model numbers sit in the paper body, not the abstract). One limit stands: it doesn’t re-test the consolidation remedy — no 2026 study does.
• “Regression Accumulation in Multi-Turn LLM Programming Conversations” (Huang et al., 2 July 2026) — the same decay in a real coding setting: across six models (GPT-4o, DeepSeek-V3 and Llama-3.1-8B among them — capable models, though none of 2026’s frontier assistants), “40% to 73% of tasks lose previously correct behavior over the full conversation.” Its mitigation tests all stayed in-thread — recap summaries helped modestly, a verification-and-rollback gate helped most — and restarting clean wasn’t tested. The modest recap result is consistent with the driver’s finding that in-thread repetition recovers only a fraction of the loss.
• “Found in Conversation” (Chen, Wu & Leskovec, Stanford, 23 May 2026) — an independent lab confirms the gap is real and large, recovering “at least 92% of single-turn performance” with a model-side self-distillation fix. The boundary worth noting: that’s a fix for model builders, not a habit you can apply — which is exactly why the user-side move is to restart clean.

Editor’s note

Many of you will already know this and be doing it. This module is aimed squarely at those who aren’t. It’s easy to get attached to a session, and we’ve all found ourselves deep in correction territory, trying to push through confused output. With a junior new starter that instinct is the right one, because a junior learns from each correction. A lost chat doesn’t, and each new correction is added to the chain that derailed it. So cut your losses, write the brief once, properly, and take it to a fresh session.