Optimal Forgetting: Information as Congestible Capital

Emiliano S. Pagnotta

Working paper, November 2026 Part of Firms and financial market structure

Draft coming soon

Abstract

An AI agent's notes record grows with its findings and is searched by its later work; past a peak, this capital congests its own use, and in the Gaussian-location instance, between the peak and the capacity, a larger record's retrieved signal is a strict Blackwell garbling of a smaller one's. With a capacity limit, a smaller post-peak record of equal positive retained precision is worth strictly more. We study costly cuts shrinking size no slower than precision. With the limit, optimal plans from nonempty self-produced records repeat such cuts under a patience condition to restore headroom. When precision raises output, each cut destroys true information productively valuable at fixed size. On a related notes record, if removal charges do not fall with the trigger and output per charged call just below the capacity falls short of the best removal loop's average, every best loop removes notes before the capacity. In live tests on two language-model agents, early removal filed more findings per charged call where resident-note success fell with record size, and fewer where it stayed near one. Implications follow for valuing information where it is used, measuring congestion, and managing memory.

Cite

Emiliano S. Pagnotta (2026). "Optimal Forgetting: Information as Congestible Capital." Working paper.