InMind: memory systems answer 14.4% of what the model gets 84% right with the same fact in context.

arXiv 2607.24368 names a failure mode every memory-augmented agent has and nobody measures: retrieval assumes the memory you need textually resembles the query, which breaks whenever world knowledge is the bridge. Stored tree-nut allergy, macaron request, zero surface overlap. On 125 expert-verified tasks across ten life domains, with the decisive memory in context the backbone hits 84.0%; when it must be retrieved, six vector, graph, and agentic memory systems top out at 14.4% despite recalling the same facts on direct demand at up to 100%. Raising embedding dimensionality 8x improves target recall and leaves the gap intact. The failure is the query-conditioned retrieval interface, not storage.