What's this about?
When a task outgrows the model's context window, it continues in a new session. This paper studies the methodological and theoretical properties of that handover — what can actually be carried across the boundary.
What exactly is "in-context learning state"?
Over a long session a model adapts to you without any training: it picks up your naming conventions, the shape of the codebase, the mistakes it already made and shouldn't repeat, the tone you prefer. None of that is in the weights — it exists only as a consequence of the conversation sitting in context. When the session ends, it evaporates. What makes this hard is that it isn't a list of facts you could write down; it's a diffuse adaptation. Summarizing captures the facts and loses most of the adaptation.
Why is this worth noticing?
- It's the most common frustration in day-to-day agent use, and it's usually treated as an engineering annoyance rather than a research question with structure worth studying.
- It connects the week's other threads: Codex's 1M context, Pi's compaction, and Baoyu's critique that compaction is lossy with no retrieval are all approaches to this same boundary.
- Framing it as "what transfers" rather than "how do we summarize" is the useful move, since it admits some of what's lost may not be recoverable through text at all.