Theodore Magnus Øen

Verified Heuristic Memory Management for Long-Context Language Models

This manuscript proposes Verified Heuristic Memory Management (VHM), a theoretical framework for governing memory in long-context language models. Candidate token spans receive dynamic eligibility weights based on salience, recurrence, utility, consistency, freshness, and evidenc...

Verified Heuristic Memory Management for Long-Context Language Models

This manuscript proposes Verified Heuristic Memory Management (VHM), a theoretical framework for governing memory in long-context language models. Candidate token spans receive dynamic eligibility weights based on salience, recurrence, utility, consistency, freshness, and evidenc...

The Magnus Matrix Tree IV: Basis-Compressed Kinetic Operators, Projected Rank-Two Transfers, and Online Selector Admission

We develop a basis-compressed operator layer for exact kinetic distance aggregation. Let x1, ..., xn be labelled coordinates in strict current order, let Cij = |xi - xj|, and fix two selector bases P in R^(n×r) and Q in R^(n×s). Rather than maintaining individual contractions, th...

Transcendental Relationalism: Framework-Indexed Determinacy, Relative Truth, and Invariant Objectivity

Transcendental Relationalism develops a framework-indexed account of determinacy, truth, and objectivity. Its central claim is that an object, identity, property, or truth-value becomes determinate only under conditions specifying a domain, language, criteria of identity, admissi...