Activation steering is being explored as a form of memory for AI agents. Instead of storing past experiences as text, the approach encodes them as steering vectors that can be retrieved later. The technique has recently been used specifically as a memory for skills.
The new paper, "What Survives When Activation Steering Is Used as Memory," tests a common assumption: that steering vectors work for skills, but facts and preferences about the user still need to be stored as text.
The abstract ends mid-sentence — "We test that…" — so the paper's findings are not available from the announcement. This digest can therefore only report the question under investigation, not the answer.