H2O Dormant Memory: Preregistered Synthetic Evaluations and a Failed Context-Selection Non-Inferiority Test for LLM Agent Swarms
Year: 2026
Type: preprint
Abstract: H2O Dormant Memory is a falsifiable prototype that keeps provenance-bearing memory units outside an active working set until typed cues reactivate a bounded subset. Six preregistered studies evaluated the approach using synthetic or repository-derived data. Early controlled tests reduced active candidates and prompt bytes while preserving quality. The final document-transformation comparison, however, failed its preregistered non-inferiority condition: submitted prompt bytes fell from 41,232 to 5,506, but quality fell from 10/12 to 6/12 because the selector omitted critical numerical evidence and a measurement-status marker. Actual billed tokens, monetary cost, provider-side memory, and energy were not measured. The result is reported as a negative finding and a design requirement for evidence-complete, reversible context selection.
Author: Sakakibara, Ryoko
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7208823114
Cites:
Cited by:
Related to: