Researchers have published a preprint on arXiv (arXiv:2603.23610v1) presenting Environment Maps, a persistent, agent-agnostic framework that consolidates heterogeneous evidence — including screen recordings and execution traces — into a structured graph to support long-horizon AI agents. The framework comprises four components: Contexts (abstracted locations), Actions (parameterized affordances), Workflows (observed trajectories), and Tacit Knowledge (domain definitions and reusable procedures). The authors describe the system as human-interpretable, editable, and incrementally refinable. Evaluation was conducted on the WebArena benchmark across five domains. Agents equipped with Environment Maps achieved a 28.2% success rate, compared to 14.2% for baseline agents limited to session-bound context and 23.3% for agents with access to raw trajectory data used to generate the maps, according to the paper. The authors frame the improvement as evidence that structured environmental representations can serve as a persistent planning foundation, mitigating hallucinations and trial-and-error behavior common in dynamic interface settings. The paper identifies robust automation of complex software workflows as an open problem in the field, noting that current large language models remain vulnerable to task failure from single missteps in dynamic interfaces. Environment Maps are positioned as a structured interface layer between the model and its environment, sourced from arXiv at https://arxiv.org/abs/2603.23610.