Deleting a record from the main database does not erase its copies from caches, logs, queues, analytics, or backups. GenAI systems can also retain prompts, outputs, embeddings, memories, and data that influenced fine-tuned model weights. Treat erasure as a lifecycle-tracing task, not a single database operation. section 5.3.3
For an erasure request or retention policy:
- Map where data was stored or transformed, including agent working memory, tool-call traces, saved outputs, long-term memory, and downstream services. section 5.3.3 section 5.6.3
- Make relevant stores searchable and purgeable. Set retention timelines, use short TTLs for working memory, and automate deletion of aged caches and records. section 5.6.3 section 2.6.2
- Check whether data entered training. Exclude it from future training if possible; if already used, assess retraining, machine unlearning, or another model-level response. section 5.6.3
An agent books an appointment
A booking task can send health data to calendar, transit, and insurance services, each of which may retain it independently. An erasure workflow must trace and purge the data across every service the agent touched. section 5.6.3
Figure 5.3 ChatGPT leaking an uploaded customer document in an unrelated query[39]. In a conversation about dentistry, ChatGPT tells the user “I'd like to show you an illustration with these steps. Shall I go ahead and select one and show it to you?” after which it surfaces an unrelated document with the results of a driving school test.

The figure illustrates how a document left in an indexing or memory layer can surface in an unrelated conversation—a failure deletion controls should prevent. section 5.3.3









