A wrong note is worse than a missing one
AI tools now remember. ChatGPT, Claude and the rest keep notes between conversations: your preferences, your projects, what they've picked up about your company. That makes them more useful. It also raises a question most teams haven't asked yet: what has it saved, and is it right?
Anthropic published research last week on how memory works for Claude, and one finding stood out to us.
A wrong note sends it in circles
In one test, Claude was playing the original Pokémon game and keeping notes as it went. It wrote down that it had finished a cave, when it hadn't. Acting on that note, it kept heading out the wrong way and landing back where it started. Three times. It never reached the next town.
Then a cleanup pass, which Anthropic calls dreaming, reread what had actually happened, caught the mistake and corrected the note. With the corrected note, Claude went back into the cave and made it through.
The note wasn't missing. It was wrong, and the AI trusted it completely. A missing note makes an AI stop and look. A wrong one makes it act, with full confidence, in the wrong direction.
Shared notes get overwritten
In another test, several AI agents worked through a set of contracts and recorded the key terms on one shared page. They kept writing over each other, and only 41 of the 57 terms survived. When each contract got its own page, with the terms quoted word for word, all 57 made it.
Long instructions get followed less
Anthropic also points to a plain warning from its own documentation: longer instruction files reduce how closely the AI follows them. The layer that applies to everything has to stay short.
What this means for your company
Swap "a cave in a video game" for "your company," and the lesson carries over. As AI tools remember more about the business, a few things start to matter.
- Keep exact wording for anything that has to be right. Product names, approved claims and legal language should be stored as written, not paraphrased.
- Give each thing one clear home. When everyone writes to the same place, things get overwritten and nobody notices what was lost.
- Keep the standing rules short. A short list of what always applies gets followed. A long one gets skimmed.
- Look at what's been saved, regularly. Notes drift, and someone has to read them and fix what's wrong.
Where Tandem fits
This is the thinking behind how Tandem works. A company's approved claims are stored word for word, each with a status. Each brand has its own guide. Standing rules are short and set by the people responsible for the brand. And the brand understanding, meaning what the AI has concluded about the company, can be read, corrected and locked, with anything new waiting for approval before it counts.
A wrong note does more damage than a missing one. The answer is visibility: someone who can see what the AI remembers about the company, and set it straight.
So if your team's AI keeps getting the same thing about your company wrong, look at what it has saved before you rewrite the request again.
Read Anthropic's research: The components of agent memory. For more on how company AI decides what's true, see our white paper, Same company, different answers.
Build boldly. Stay the course.