Finding disagreement

Contradiction flags, not verdicts

Lexical overlap plus opposite negation marks a conflict for a human. It never picks a winner.

The signal

On write, the new memory is compared with up to the 200 most recent live memories in the same space. Two conditions must both hold:

• Lexical similarity ≥ 0.55, computed as term-frequency cosine over the normalised term vectors.
• Opposite negation polarity. One text contains a negation marker and the other does not.

What a flag contains

{
  "memory_id": "mem_...",
  "similarity": 0.71,
  "resolution": "unresolved",
  "note": "lexical overlap with opposite negation polarity; review, not a verdict"
}

What it refuses to do

• It does not delete, hide or supersede the prior memory.
• It does not choose which statement is true.
• It does not change status: the new memory remains active.
• It does not claim semantic entailment. Every flag says resolution: unresolved.

Known false positives and negatives

A genuine update (“the limit was 500; it is now 800”) may not be flagged at all, because the two sentences share little vocabulary. A nuanced correction may be flagged when the negation is incidental. This is a triage signal whose only promise is that it never silently overwrites.

memory-integrity is the deeper auditor in this stack: it reads a full write history and scores corroboration. AETHER-33's inline check is deliberately cheap so that write latency stays flat.