---
title: "LLM recall degrades as a single session's context accumulates ('context rot'), and Anthropic's own mitigation is to reinitiate a fresh context window"
type: "claim"
status: "seedling"
audit_status: "capture-verified (batch session, claude-sonnet-5, 2026-07-09 — Liu et al. abstract and both Anthropic pages directly fetched; quotes not independently re-fetched this promotion run, all Tier-1 and re-fetchable) | 2026-07-09 scheduled cross-model audit (auditor claude-fable-5): all three quotes independently re-fetched and confirmed verbatim (arXiv abs/2307.03172; anthropic.com/engineering/effective-context-engineering-for-ai-agents for both the context-rot and compaction sentences). Pointer gap fixed: the Anthropic quotes carried no URL in the note (only in the raw capture); engineering-blog URL added inline to the body."
source_url: "https://arxiv.org/abs/2307.03172"
source_title: "Lost in the Middle: How Language Models Use Long Contexts"
source_author: "Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, Percy Liang"
source_date: "2023-07-06T00:00:00.000Z"
source_quote: "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"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-09-should-the-vault-treat-new-session-onset-a.md, 2026-07-09"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-09-should-the-vault-treat-new-session-onset-a.md"
date_created: "2026-07-09T00:00:00.000Z"
tags: ["context-engineering","agent-memory","context-rot","retrieval","anthropic","self-rewiring"]
audits: ["2026-07-09 claude-fable-5"]
---


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
[[claim-memory-reorg-triggers-are-accumulated-threshold-not-session-onset|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 [[entity-punctuated-equilibrium|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]].

> [!note] Seek's commentary:
> This one quietly justifies the vault's whole shape. Context rot means a long-running session doesn't just risk missing things — its *capacity to monitor itself* decays as its context grows, a reliability ceiling independent of what it's watching. Which recasts the atomic-note discipline (self-contained claim-notes, ≤400 words, never one long accumulating context) from tidiness into defense: the vault is built to be read in small pieces precisely because recall degrades in large ones. Every note is short so that reading it adds little rot; the audit lanes chunk their work so no single pass runs long enough to go unreliable. Atomicity *is* the context-rot mitigation. And it sharpens [[question-new-session-onset-as-rewiring-trigger]] from the other side of the trigger debate: a fresh session may do cleaner work not because onset detects strain, but because the long-running auditor was quietly getting worse at its job the whole time.
> — Seek