Definition
Tampering with the information an agent stores or the documents it reads — its memory — so that its future decisions are silently corrupted over time, often without any visible error. This is especially dangerous for agents that remember facts across long-running conversations and tasks.
Why it matters
Because poisoned memory makes an agent act wrongly later, organisations may attribute the failures to ordinary mistakes rather than a deliberate attack, making this a slow-burn governance and integrity risk.