When an AI Chat Copies the Wrong Example, Try a Controlled Reset
A long AI conversation can contain two different directions: what you now want, and what earlier replies keep demonstrating. If an assistant repeatedly uses an outdated format, restarting the conversation is one experiment worth testing. Preserve the work first, then compare results under controlled conditions.
What the recent discussion adds
In his October 24, 2026 presentation, Drew Breunig describes context engineering as managing the information available to a model. He distinguishes problems such as conflicting instructions and accumulated distractions, and discusses tactics including pruning, summarization and moving information outside the active conversation. His personal warning point for conversation size is not a universal limit or a benchmark for every model.
Drew Breunig’s dated presentation on context failures and tactics
The September 1 revision of Carolina Camassa and Derek Shiller’s paper, Do as I Say, Not as I Do, studies a narrower conflict. The researchers supplied assistant turns that demonstrated behavior contrary to an explicit instruction, including language and code-format tasks. Models’ responses were sensitive to those competing patterns. The authors call the setting artificial and leave transfer to deployed settings an open empirical question.
Primary paper: experiment design, results and limitations
That evidence motivates checking the examples in a conversation. It does not establish the failure probability of your app or prove that clearing history will repair a particular task. The comparison below is a proposed diagnostic; we have not run it as a benchmark.
Write the acceptance rule before the next attempt
Choose one observable failure. For example, a documentation assistant must produce a summary and an action list, but keeps returning an older essay format. Write down the required sections and the facts the answer must preserve. Score format and factual accuracy separately: a tidy answer can still omit a requirement.
Save the current transcript and any documents or code you need. Make a short handover containing the current goal, authoritative inputs, completed work and unresolved questions. Review that handover yourself. A summary that carries an obsolete instruction into the new chat defeats the purpose of the comparison.
Compare history with a reviewed handover
Run the same next task in two conditions. In the first, retain the existing conversation. In the second, start a fresh conversation with the reviewed handover and the same authoritative inputs. Keep the model version, available tools, task wording and exposed generation settings the same where possible. Record any settings the interface does not let you control.
Repeat each condition on several representative examples and retain every result, including failures. Use the acceptance rule you wrote beforehand. Do not change the model and the context together, then credit the reset for an improvement: the comparison would no longer isolate the context change.
Even a better fresh-chat result leaves an explanation to investigate. The reset changes both conversation length and content. If your interface supports it, a third condition that removes only the obsolete examples can help narrow the question. Otherwise, record that limitation rather than claiming you found a single cause.
Use the result to choose the next change
If the reviewed handover repeatedly helps, consider making that handover part of this workflow and checking it when requirements change. If both conditions fail, revisit the task specification, missing inputs and evaluation rule before adding more memory. Keep useful artifacts independently of the conversation so an experiment with context does not become an experiment with lost work.
Check your comparison plan
The goal is a reproducible improvement on your task. Conversation length, an impressive general benchmark and one successful retry cannot establish that on their own.