Nate B Jones states that long-running agent sessions inevitably fill up context windows, and all compression strategies are inherently lossy.

“This claim accurately describes a fundamental limitation of current large language model (LLM) architecture. LLMs operate with a fixed-size "context window," which limits the amount of information they can process at once. In long-running agentic sessions, the history of interactions and observations eventually exceeds this limit. Techniques to manage this overflow, such as summarization, truncation, or retrieval-augmented generation (RAG), are inherently lossy because they either condense or selectively omit information to fit within the window, meaning the model does not have access to the complete, unabridged history at all times. Context: This is a core technical challenge in building stateful, long-term AI agents. The problem of context window management is a major area of research and engineering in the field of applied AI, as it directly impacts application performance, cost, and user experience.”