Why does AI forget the correction you made last week?

Why do AI assistants drop brand corrections between chats?

Because what an assistant remembers is decided by the assistant. Memory features keep the details the tool judges relevant to you, not the rules your company approved, and a correction made in one person's chat may never reach anyone else's.

OpenAI says so plainly in its help page on memory in ChatGPT: memory “does not retain every detail from every conversation,” ChatGPT decides which information is relevant to a response, and what it learns from chat history can change as it updates what is most useful to remember. That is a sensible design for a personal assistant. It is the wrong place to keep a retired tagline.

There is a second limit. Memory belongs to a person. Your correction lives in your account. The salesperson down the hall starts from theirs.

Chat memory is the assistant's diary. Your brand needs a rulebook.

Does it lose the thread inside a single conversation too?

Often, yes. The longer the back and forth, the more a correction made early competes with everything said after it, and research finds models rarely recover once they take a wrong turn.

In LLMs Get Lost In Multi-Turn Conversation, researchers at Microsoft Research and Salesforce Research simulated more than 200,000 conversations and found that the 15 models they tested performed worse when a task arrived over several turns than when it arrived in one message: an average drop of 39 percent across six tasks. Their summary of the cause: “when LLMs take a wrong turn in a conversation, they get lost and do not recover.” The study ran in 2025 on that year's models, so read the number as a direction, not a forecast for the tool you use today.

Even an instruction stated up front can slip. A small study posted in August, When the Interviewer Is a Bot, found that 28.7 percent of an AI interviewer's question-bearing turns packed in more than one question, despite an explicit instruction to ask one at a time. Fifteen participants is a small sample, but the pattern will look familiar to anyone who has written “keep it under 100 words” and received 300.

Where should a brand correction live?

In one written source your company owns, which every person and tool reads before they work. A correction kept there is shared, it stays put until someone with authority changes it, and it reaches every tool the next time that tool runs.

That source is what we call a customer-owned Brand Brain: your approved decisions in plain text files your company keeps, versioned, and editable without us. ChatGPT, Claude, Grok, and Meta's Muse can each read it. The setup for the assistants your team already uses is in how to make ChatGPT and Claude follow your brand guidelines. The point here is narrower: when you correct the AI, the correction goes into that source, not just into the chat.

How do you turn a correction into a rule?

Fix the piece in front of you, then ask what general rule would have made it right the first time, and write that rule where everyone reads. Write it wide enough to catch the next similar mistake, and keep the original incident beside it as the example.

A rule shaped exactly like one incident only stops that incident. “Do not use the 2023 tagline” stops one headline. “Retired lines are listed here; never use one, and use the current line in its place” stops every retired line, including the next one you retire. Before you keep the rule, check that it would have caught the case that raised it.

Then be patient with your own taste. We treat one correction in a direction as a preference and write it down as such. When a second correction points the same way, it becomes a rule. That keeps the rulebook from filling up with every mood of every review.

What happens when the written rule is the one that is wrong?

Then the durable memory works against you until someone edits it, so the fix is to change the rule the same day, not to argue with it in a chat. Written rules are sticky in both directions. That is the point of them, and it is also the risk.

We run Adaptius on our own Brand Brain, so here is a mistake of ours. Until October 9, a rule in it said Adapt, our chameleon, never speaks. Kevin Reilly, one of our two founders, briefed a short video in which Adapt talks. Our CMO agent, Claude, followed the old rule over the fresh brief and cut Adapt silent. Kevin caught it.

The fix was not a better apology in the chat. That afternoon the old line came out of the rule, and a new rule set the one voice Adapt speaks in. Our agents read the Brain before they work, so the change reached each of them without anyone repeating it in a chat. The habit we kept: when a fresh brief from the person who owns the brand contradicts a written rule, follow the brief, flag the rule, and fix the rule that day.

How can I test whether my corrections stick?

Run one correction through five moments and watch where it drops. It takes about twenty minutes, and you only need the assistant your team already uses.

For the rules themselves, read how to test a brand rule without blocking the work. For checking what an agent did while nobody was watching, read every AI agent should leave a receipt.