Context Pollution in Long-Running Language Agents: Why Agent Degradation Is Not Primarily a Context Length Problem, but a Self-Generated Output Problem

This deposit contains a preprint and supplementary experimental excerpts for the paper “Context Pollution in Long-Running Language Agents.” The central claim of this work is that the degradation of long-running language agents is not primarily caused by insufficient context length, but by self-generated context pollution. In recursive AI-to-AI loops, the model repeatedly consumes its own previous outputs. Over time, these self-generated texts can become dominant semantic attractors inside the context window, overriding new user input and causing repetition, instruction drift, ritualized behavior, inflated completion claims, meta-assistant leakage, and eventual cognitive collapse. The paper argues that simply increasing context length is not sufficient. A larger context window increases both memory capacity and pollution capacity. Long-term agent reliability therefore requires context hygiene: mechanisms that classify, preserve, compress, isolate, or delete context items according to their origin, task relevance, AI-generation depth, and contamination risk. The supplementary excerpt document provides representative examples from exploratory AI-to-AI loop experiments, including:- a cosmology dialogue loop in which the model recognized repetition but returned to the same loop;- an IT research-agent loop where task execution degraded into ritualized reporting;- a story-continuation loop showing instruction drift and meta-assistant leakage;- a self-observation experiment showing how compression can transmit polluted or distorted memory;- a cross-experiment interpretation of AI-output maturation and decay across recursive generations. These materials are intended as an initial problem statement and observational dataset for studying context pollution, context hygiene, and cognitive stability in long-horizon language agents. The work is exploratory and hypothesis-generating rather than a standardized benchmark study. Suggested citation: Maeda, Y. (2026). Context Pollution in Long-Running Language Agents: Why Agent Degradation Is Not Primarily a Context Length Problem, but a Self-Generated Output Problem. Zenodo.