Can LLMs read long documents like humans do?

LLMs are limited not only by an explicit context window but by degrading performance on long inputs well before that limit. ReadAgent's premise is that humans read differently: exact wording is forgotten quickly while gist — the substance irrespective of exact words — persists, and reading is interactive (we look back when we need a detail). It implements this as a simple prompting system that (1) decides what content to store together as a memory episode, (2) compresses each episode into a short gist memory, and (3) looks up the original passages only when a task requires the details. This extends effective context 3–20× and outperforms retrieval baselines on QuALITY, NarrativeQA, and QMSum.

The keeper is that the LLM can generate broadly useful gist memories before knowing the task — compression need not be query-conditioned to be useful — and can then reason interactively over those gists to decide what to retrieve. Gist-first-then-lookup is a different long-context strategy than either stuffing the window or pure retrieval.

This is the gist-compression member of the vault's long-context/memory cluster. It shares the compress-then-act move with Can agents compress their own memory without losing critical details? and the bounded-state philosophy of Can agents fail from weak memory control rather than missing knowledge? — but applied to reading documents rather than managing agent state, and beating retrieval rather than replacing it.