When One Source Returns: A Preregistered Behavioral Study of Intra-Agent Evidence Recycling

Abstract

Large language model agents increasingly retain summaries, reviews, and execution traces derived from earlier observations. Such records may later re-enter context alongside the evidence from which they were produced, creating a risk that one epistemic source acquires additional behavioral influence. We call this possibility intra-agent evidence recycling. This preregistered, fixed-N behavioral study used 32 balanced fictional binary-choice items and qwen3.5-4b under a frozen local inference configuration. All 168 planned trajectories were valid and all prespecified validity gates passed. Passive repetition produced initial-claim retention in 22 of 32 items, compared with 0 of 32 under length-matched neutral memory. The paired risk difference was 0.6875, with a Holm-adjusted exact McNemar p-value of approximately 9.54 × 10⁻⁷. The preregistered lineage-mitigation contrast was not supported: 2 of 32 versus 0 of 32, with a Holm-adjusted p-value of 0.50. The inference is limited to the tested model, local inference configuration, and fictional binary-claim task family. This preprint has not been peer reviewed. Code, preregistration, raw results, analysis scripts, and integrity manifests are available in the associated public software record and GitHub repository.