State-Drift
A failure mode in AI agent memory systems in which beliefs or facts stored in the system have ceased to be true over time, yet the agent still retrieves and acts on the outdated records with full confidence, producing misalignment along the temporal dimension.
Timeline (last 90 days)
State-drift describes a situation where an Agent's memory system recorded information at some point in time, and after the facts in the world changed, the system still retrieves and acts on the old record with full confidence, while the retrieval pipeline itself is not at fault.
State-drift is different from data drift or concept drift, which focus on changes in the model's input distribution or input-output mapping and can be detected with observability tools; state-drift concerns the expiration of stored beliefs, which can fail even when the input is entirely in-distribution.
When the world changes on its own, an outdated belief is a reliability bug; when someone deliberately changes the world, it becomes a vulnerability, and the mechanism of the two is exactly the same.
No matter how strong an Agent's memory capability is, if it cannot determine which memories have become invalid, then every retrieval in a dynamic environment may plant the seed of an erroneous action.
All Facts (4)
State-drift describes a situation where an Agent's memory system recorded information at some point in time, and after the facts in the world changed, the system still retrieves and acts on the old record with full confidence, while the retrieval pipeline itself is not at fault.
50%UnverifiedState-drift is different from data drift or concept drift, which focus on changes in the model's input distribution or input-output mapping and can be detected with observability tools; state-drift concerns the expiration of stored beliefs, which can fail even when the input is entirely in-distribution.
50%UnverifiedWhen the world changes on its own, an outdated belief is a reliability bug; when someone deliberately changes the world, it becomes a vulnerability, and the mechanism of the two is exactly the same.
50%UnverifiedNo matter how strong an Agent's memory capability is, if it cannot determine which memories have become invalid, then every retrieval in a dynamic environment may plant the seed of an erroneous action.
50%