T3. Temporal belief revision

A plain vector store has no notion of a decision changing. When a fact is updated, the old record still sits in the index and can be retrieved as if it were current. This is the failure that produces confidently stale answers.

Anamnesis treats truth as temporal. We tested it directly:

1. A decision is recorded (an earlier value for a setting).
2. A month later the decision changes, in a note that references the earlier one.
3. A contradiction sweep runs.

Result: every check passed. The sweep detects the supersession automatically and retires the old record. From that point, what the system injects as current context contains only the new value; the superseded record is inhibited. Nothing is deleted, so the earlier value is preserved for audit and for point-in-time ("what did we believe on this date") history.

This is the counterpart to the scale benchmark. Trust ranking (see T2) resolves contradictions that differ in provenance; belief revision resolves contradictions that differ in time. Together they keep the current, authoritative answer in front of the model as a memory pool grows and evolves.