What's this about?
EngramEdit changes factual associations stored in conditional n-gram memory without updating the main transformer. It aligns multiple ways of expressing the new fact while protecting memory entries shared with unrelated facts.
What exactly is conditional memory?
It is a lookup table activated by patterns in the input, sitting beside the neural network's general computation. Moving some facts there is like keeping an editable reference drawer next to a fixed reasoning engine.
Why does this matter?
Editable externalised facts could make corrections cheaper and reduce collateral damage from model editing. Near-perfect success and the reported reasoning gains come from the authors' tests and need broader-scale validation.