Skip to content
InfoResearchPreprintLLM-specific

Forget the Lineage: Forward-Secure, Lineage-Complete Cryptographic Erasure for LLM-Agent Memory

Published
Record updated
View JSON

Summary

LLM agents store long-term memories that the model writes from user input, so deleting a message leaves derived memories, storage residue and backups intact. The authors present LinEra, which encrypts each memory under keys derived from its parents along the derivation graph, so one puncture of a source key makes all descendants unreadable. Across ten LoCoMo conversations, an attacker holding all storage and current keys recovered 70 of 71 forgotten secrets after logical deletion, and none with LinEra.