Gardado en:
| Autor Principal: | |
|---|---|
| Formato: | Recurso digital |
| Idioma: | inglés |
| Publicado: |
Zenodo
2026
|
| Subjects: | |
| Acceso en liña: | https://doi.org/10.5281/zenodo.19353247 |
| Tags: |
Engadir etiqueta
Sen Etiquetas, Sexa o primeiro en etiquetar este rexistro!
|
Table of Contents:
- <p>This paper proposes an architecture for LLM context management that separates commands and data into distinct memory spaces (L1 command, L2 active data, L3 reference storage) with default-deny triage: all incoming content enters the data tier by default and must pass through evaluation and summarisation gates before promotion to the command tier. The architecture addresses both prompt injection vulnerability and context degradation as consequences of the same architectural feature: the undifferentiated context window. Two triage modes are specified: a security-first mode (L3-by-default with two-gate promotion) and a performance-optimised mode (platform triage).<br>The paper identifies a verification inversion where unmanaged context structurally penalises fact-checking, reframes the entanglement problem from weight-level disentanglement to contextual mode-setting, and grounds the gate mechanism in peer-reviewed empirical work on task-frame shift and adversarial robustness. Native Memory Paper 2 in the series. Builds on Strategic Forgetting (Paper 1) and findings from the Confidence Curriculum series.</p>