Belief Updating

Changing what a system holds true when new evidence arrives: detecting that two claims disagree, deciding which survives, and recording the change without destroying what came before.

Why it matters in a memory layer

It is the mechanism with no counterpart in retrieval, and the one every characteristic memory bug traces back to. A corpus does not change its mind; a memory layer does nothing else.

The division of labour is what makes it safe. Detecting that two beliefs conflict is a language judgement, and a model does it well. Deciding which wins is policy, and a model doing it produces different answers on different runs with no record of why — which is how a correct belief gets silently overwritten. Model detects, rules arbitrate.

The other discipline is that updating is not deleting. A retired belief keeps its content and gains an invalid_at, so "what do I believe now" and "what did I believe in March" are both answerable — see supersession.