Keeping beliefs current

Supersede, don't delete

When a fact is corrected, ENGRAM derives a new memory that supersedes the prior one, stamps the old belief with the moment it stopped being true, and links the two in a version chain. Recall returns only the head of the chain — the current belief — while every retired version stays in the graph. Because each version carries validity bounds, ENGRAM can answer point-in-time questions: what an agent believed about something on a particular date, from the same store.

Recall returns the head. Retired beliefs are kept, with validity bounds, for audit and point-in-time recall.

Catching contradictions

Two triggers feed the same path. The inferred one fires when a new belief lands: ENGRAM compares it against existing memories and puts a near-neighbour to an entailment check that decides whether the new belief supports, contradicts, or is unrelated to the old — only a contradiction proceeds. The deterministic one needs no inference at all: when a revised article supersedes an older one, the memory imprinted from the old article is superseded by the memory from the new. The version bump is authoritative.

Governed by confidence

How a contradiction is resolved depends on how sure ENGRAM is, across three fail-closed tiers. High-confidence updates are auto-applied, retiring the older belief in place. The uncertain middle band is put to a system judge — a separate, system-side model, not the one you configured for your own agents — which rules on the supersession. Anything genuinely uncertain is sent to a human, surfacing in a review inbox where the two beliefs sit side by side to be kept or retired. A missing verdict defers; it never auto-applies.

The full design — validity bounds, the entailment check, and the review flow — is told on the blog.