Deleting a message no longer deletes the memory: LinEra voids an entire derived chain with one forget request, and a thief holding every file reads nothing
Synopsis
An LLM agent's long-term memory is mostly written by the model itself, so deleting the original message leaves summaries, facts and embeddings in storage and backups. LinEra ties each memory's key to the keys of its parents, so forgetting a source is a single puncture of a puncturable pseudorandom function and every descendant becomes undecryptable at once. On LoCoMo conversations written by Qwen2.5-7B, an attacker holding every file and the current keys recovers 70 of 71 forgotten secrets after logical deletion and 67 with per-object or epoch keys, and none with LinEra.
Interpretation
Forgetting one user input can now invalidate every memory derived from it at the same time, without first locating those memories. Logical deletion trusts the store to behave after the fact, and per-object or epoch keys erase only the named object, leaving the model's summaries and facts readable. On LoCoMo conversations whose memory was written by Qwen2.5-7B, an attacker holding every file and the current keys recovers 70 of 71 forgotten secrets after logical deletion, 67 with per-object or epoch keys, and none with LinEra.
The mechanism makes each memory's key derive from the keys of the parent memories the writer saw, so forgetting becomes a key operation. Source keys come from a puncturable pseudorandom function, and puncturing one point leaves that source and all its descendants without a decryptable key, wherever their ciphertexts live. Integrated into mem0 2.2.1 through an adapter of about 140 lines, forgetting two messages containing secrets erased 12 objects, and the attacker found no secret in decryptable records or in raw bytes.
The price is a larger forgetting radius, set by how many source turns enter the writer's context. A session-level writer has 181.5 source ancestors on average, and forgetting 1% of turns erases 73.0% of derived memories; four-turn windows cut fan-in to 3.8 and the radius to 3.8%. On the same LoCoMo conversations, the four-turn window writer keeps answer quality at 30.7 when 1% is forgotten, while the session writer falls from 27.8 to 13.8 and returns to 27.9 after re-derivation.
Perspective
The experiments use open models of at most 7B parameters: Qwen2.5-1.5B writes memory on five LoCoMo conversations, Llama-3.2-3B rechecks two conversations, and Qwen2.5-7B replays all ten conversations and 100 LongMemEval-S histories; answer quality is compared only across conditions computed on one machine.
The shape of the lineage graph, object sizes and creation times stay in clear, so after forgetting an attacker learns that some object with children existed but not its content. The scheme does not protect model weights, RAM at the moment of forgetting, or copies made before the request; if the same fact appears in another message, forgetting one message does not forget the fact.
