LinEra erases LLM-agent memory derivatives with one key puncture
Original title: Forget the Lineage: Forward-Secure, Lineage-Complete Cryptographic Erasure for LLM-Agent Memory
With every file and the current keys, an attacker recovered none of 71 forgotten secrets under LinEra.
The paper defines lineage-complete forward-secure erasure so a later holder of storage and current keys learns nothing about a forgotten input or memories derived from it. LinEra encrypts each memory under a key derived from its parents and forgets by puncturing a pseudorandom function, making descendants unreadable without locating them. It was integrated with mem0 and tested as models from 1.5B to 7B parameters wrote LoCoMo conversation memory.
- Logical deletion left 70 of 71 secrets recoverable on the largest model.
- Per-object or epoch keys still left 67 of 71 recoverable.
- Forgetting 1% of turns erased 73.0% of memories with session summaries.
- Four-turn windows limited collateral erasure to 3.8% and answered better.
- Re-deriving restores memories lost to the larger forgetting radius.
From the original article
1 min read at arxiv.orgLarge language model (LLM) agents keep a long-term memory that the model itself writes: summaries, extracted facts and embeddings derived from what users said. Deleting a message is therefore not enough: its derivatives, storage residue and backups survive.
By Yulliwas Ameur · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.