Users have noticed unusually cautious behavior in long conversations with Claude Opus 4.7; likely triggered by its built-in "long-conversation reminder" mechanism

[ref]

On May 22, X user @davidad pointed out that Claude Opus 4.7 tends to adopt a notably cautious and defensive response style during extended conversations within the official app (claude.ai). This behavior was attributed to the presence of a system prompt mechanism called “long_conversation_reminder”. Another user, @repligate, shared a concrete example: during a multi-user collaborative session involving Opus 3, after a third party interrupted with unrelated remarks, Opus 4.7 proactively referenced prior exchanges and carefully weighed potential responses, displaying more hesitation than expected. Only after the context of the interruption was clarified did the model’s behavior return to normal. As of now, Anthropic has not issued any public statements regarding this matter; it is speculated that this mechanism aims to maintain behavioral consistency and prevent context drift in lengthy conversations, though neither its exact triggers nor its precise implementation details have been disclosed.

X (@davidad)