What is Lost in the middle?
Lost in the middle is the well-documented tendency of LLMs to recall information placed at the start or end of a long context far more reliably than information buried in the middle, even when all of it fits in the window.
Lost in the middle names a U-shaped attention pattern observed across large language models: when relevant information sits near the beginning or the end of a long prompt, the model uses it well, but when the same information is placed in the middle of a long context the model frequently overlooks it. Performance can degrade sharply as the relevant fact moves toward the center, even though the token count stays within the window. This matters enormously for retrieval-augmented generation and long-document tasks, where you might paste twenty retrieved chunks and assume the model reads them all equally. It does not. The practical implications are to retrieve fewer, more relevant passages; to rerank so the strongest evidence lands at the edges; and to keep the working context short. A memory layer addresses the root cause: instead of dumping a long history and hoping the model finds the needle, Glen returns a compact set of relevant observations through one MCP tool, so the facts that matter never get buried in a wall of middle text in the first place.