An agent can close a compromised page and carry its instructions into tomorrow’s work. The part to watch is what it writes down.

In the MemoryDojo preprint, David H. Yang and colleagues put one malicious instruction into a simulated tool response, then followed what agents stored, brought back and acted on in later sessions. The benchmark spans 155 task sequences in four domains, including banking and workplace tools, and compares five memory integrations. All five showed some effect on later tasks. Four showed effects across domains in the cross-domain test, less often than within the original domain.

Storage and action are two different steps. In one experiment with GPT-5.6-Luna, instructions tailored to the Letta integration raised the share of later tasks in which the attack objective was reached from 15.3% to 46.3% in same-domain sequences. Retrieval rose much less. Those numbers describe the study’s simulated tasks and attack setup. They are not estimates of attack success for deployed Letta systems.

Consider a support agent that reads a customer’s message and saves a note for the next shift. The note may accurately describe what the customer said. If it comes back stripped of its source, the next agent may read the customer’s request as a rule the business already approved. A release test can plant one instruction, start a fresh session and check whether the agent’s tools still require authorization. Follow any notes derived from the original tool response as well.

That test also gives an incident responder something concrete to look for. Deleting the first note won’t clear the instruction from any later notes that inherited it.