Activation memory used stored facts in at most 15% of tests
A stored dislike can also come back as a like, unless the writer subtracts the opposite.
Claimed, not confirmed
Researchers tested a memory that stores the change of an experience in a model as a steering vector. The vector is added back in a new session when the topic matches. The memory leaves 98% of unrelated replies unchanged. But the model used stored facts in at most 15% of tests. With the facts in the text, the rate was 96% to 98%. The tests used Qwen3.5-2B and 9B.
Sources
Posted
