How to detect poisoned memories in AI agents after the damage is done?

LLM agents that store and retrieve past interactions are vulnerable to adversaries who sneak malicious records into memory. MemAudit detects which stored memories caused harmful behavior after the fact by measuring each memory's causal influence on outputs and flagging structurally odd entries. Tested against realistic memory-poisoning attacks, it eliminated success rates that previously reached 83% in reasoning tasks.