LLM recall degrades as a single session's context accumulates ('context rot'), and Anthropic's own mitigation is to reinitiate a fresh context window

Liu et al.'s "Lost in the Middle" (2023, arXiv:2307.03172; TACL 2024) established that retrieval from a long context is position-dependent: "performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models." Anthropic's engineering blog (anthropic.com/engineering/effective-context-engineering-for-ai-agents) names the accumulation-driven version of this "context rot": "as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases." Its recommended countermeasure is literally a fresh-session move — compaction "is the practice of taking a conversation nearing the context window limit, summarizing its contents, and reinitiating a new context window with the summary."

The argument this supports is distinct from the accumulated-threshold trigger finding: it is not that onset detects something strain-detection misses, but that a system's own capacity to monitor strain reliably may itself decay as one session's context grows, independent of whether any watched threshold has been crossed. If that generalizes to a note-tending agent, it favors bounding how much a single session self-audits before restarting — but it does not by itself make the restart the rewiring trigger; it establishes only that long-running self-audit inside one accumulating context has a known reliability ceiling.

This is the mechanism most directly relevant to question-new-session-onset-as-rewiring-trigger: it is a concrete, sourced reason a fresh session might do cleaner work than a long-running one, separate from any punctuated-equilibrium analogy. It also grounds why the vault keeps atomic, self-contained claim-notes rather than long accumulating context — cf. the retrieval-load concerns in claim-contextual-retrieval-failure-rate-figures and question-embedding-layer-threshold-crossed. See moc-schema-change-and-restructuring.