Agentic Memory Should Localize Compression

Long-horizon LLM agents require memory, but unbounded storage is unusable at inference time, making compression unavoidable. In continual deployment, compression becomes repeated updates to accessible state and can induce behavioural drift on previously supported queries. We formalize this as interference: expected divergence between the agent’s policies before and after an update. Our position is that stability is governed by retrieval–update overlap; modular designs minimize overlap and thus localize update effects. Under routing stability, expected interference is controlled by the probability that updated modules are retrieved.