Abstract
<title>Abstract</title> <p>Conversational agents are increasingly expected to remember what a user told them across sessions, including the case where the user corrects something said earlier. A natural proposal is to exploit the geometry of the embedding space: rather than overwriting a memory, deposit the revised version near its predecessor, so that proximity encodes the relation between versions. We test that proposal directly, in a non-parametric index over a frozen encoder, with every prediction registered and hash-frozen before the corresponding data existed. The result is a consistent negative with an identified mechanism. With a fixed displacement step, the revision cluster performs an unbounded random walk away from its anchor: cosine to the query falls from 0.811 to 0.324 after four revisions and to -0.139 after six, and coverage collapses to 0.000 from K >= 4. Bounding the displacement repairs the drift exactly as intended (the cosine becomes flat at 0.8177, slope +0.0000 per revision) and the mechanism still loses, by -0.0864 (95% CI [-0.1014, -0.0714]) at K = 8. The reason is a constant toll of approximately 0.036 cosine: the embedding of the revision's own text is already optimally placed, because it contains the entity the query mentions, so any artificial displacement can only move it away. That last clause is a property of the generator, not of corrections. Real conversational corrections are elliptical: "no, it's Beto" does not name the entity. We therefore ran an adversarial third experiment in the regime that should favour geometry, against the honest baseline (coreference hydration, as deployed by production memory systems), sweeping its failure rate tau. Geometry only wins above tau* = 0.45, a coreference resolver worse than a coin flip, so the negative extends rather than reverses. The useful positive finding is independent of the mechanism under test: raw elliptical corrections are unrecoverable, scoring 0.0000 recall at k = 5 across all ten seeds, not low but never. In a separate check, their top-1 rank is likewise indistinguishable from chance. A memory system that archives conversational turns without resolving coreference loses 100% of corrections, and loses them silently: the index always returns something, just never the right thing.</p>