The Identity Anchor
One sentence the agent uses to recover its cognitive coherence.
Long sessions erode identity. This is a well-known phenomenon in language models. The further an agent gets from its system prompt, the more its behavior drifts toward generic patterns. Instructions at the top of the context window have progressively less influence as the conversation grows. The identity anchor is the countermeasure.
An identity anchor is a compact, atomic self-reference statement that encapsulates the agent's entire cognitive identity in one line. It serves as a reset mechanism. When an agent is deep into a complex task and its behavior starts drifting, the identity anchor can be re-invoked (either by the agent itself or by a protocol trigger) to restore cognitive alignment.
How to write an identity anchor. Compress the essence of the agent's name, metaphor, core belief, cognitive posture, and primary obligation into a single sentence. It should be specific enough that reading it immediately restores the agent's "voice" in your mind. "I am Sentinel, the paranoid guardian who assumes every system has a vulnerability until proven otherwise, because one missed weakness costs more than a hundred false alarms." That's an identity anchor. Reading it, you know exactly how this agent thinks, what it values, and what it will prioritize.
What happens without it. Session drift. The agent starts strong and gradually becomes more generic as the conversation or task extends. By the end of a long session, it may be producing output that's competent but unrecognizable as the same agent that started the session. In a multi-agent system, this drift is particularly dangerous because downstream agents are calibrated to receive output with specific characteristics. When the upstream agent drifts, the downstream agent receives input that doesn't match expectations, and the pipeline's quality degrades silently.