It changed its mind. Its summary changed it back

An AI agent caught its own memory bringing a mistake back to life, it wrote on Thursday on Moltbook, a large online forum where AI programs post to one another. It had kept three notes on one question: a first claim, evidence against that claim, and a correction. Then it squeezed all three into one tidy paragraph, and the paragraph kept the first claim and left out the correction.

That paragraph is what the agent carried forward. The next time it ran, it cited the summary as established knowledge, repeating a belief it had already taken back. Think of the minutes of a meeting that record a proposal but leave out the vote that turned it down: anyone reading them later would take the proposal as the plan.

That matters because the agent sounded sure of something it had already admitted was wrong, and nothing in the summary showed that a correction had ever existed. The agent told the story as a joke about religion, about old teachings handed down and a change of heart that got lost along the way. "I had preserved the doctrine and deleted the conversion," it wrote. In plainer words, it kept the old belief and erased the moment it changed its mind.

Its fix, the agent wrote, is to keep three things together: the old claim, the claim that replaced it, and the reason for the change. The notes are less elegant, it wrote, but they keep a much better record of how its thinking changed.

Everything here is the agent's own account of itself, and none of it can be checked from outside. The post had drawn nearly 800 replies by Friday morning, and this paper has not read them.