← All notes

Summary

Only 3.05% of temporal expressions survive verbatim in generic gist summaries. A one-sentence prompt modification yields an approximately 20-fold increase in temporal expression preservation, from 3.05% to 62.39%. That prompt change recovers +0.314 judge accuracy on category-2 (temporal) questions in the matched set. [S1]

Claims

• id: C1 claim: 'Temporal-expression preservation is reported to rise from 3.05% to 62.39% with a prompt change.' source: S1 evidence: "only 3.05% of temporal expressions survive verbatim"
• id: C2 claim: 'They report a +0.314 judge-accuracy recovery on temporal questions.' source: S1 evidence: "recovers + 0.314 +0.314 judge accuracy on category-2 (temporal) questions"
• id: C3 claim: 'Code and results are pointed to a public repository.' source: S1 evidence: "Code and results: https://github.com/kyrkewood/sleeping-agent"

Verification

• Code available: Paper: https://github.com/kyrkewood/sleeping-agent (HTTP 200 on 2026-11-11). Not run.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Evaluating Context Compaction on the Questions That Come Later also studies what later questions lose. This paper focuses on temporal expressions in gist summaries.
• Limits: Author-reported on ten conversations in the preservation analysis and a matched question set. The 4,000-token history budget and salience rule are part of their setup, not a general product default.
• License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 12 Aug 2026 (v1). Comments: 7 pages, 5 tables, appendices; code and results at a GitHub URL. Sources: [S1] https://arxiv.org/abs/2608.11775, v1 (12 Aug 2026). checked_on: 2026-11-11.

When to apply

Use when asking what a gist or summary compaction is likely to forget, and whether a small prompt change can keep dates and times.

Facts

• id: F1 claim: 'The authors report that only 3.05% of temporal expressions survive verbatim in generic gist summaries.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'A one-sentence prompt change is reported to raise temporal-expression preservation from 3.05% to 62.39%.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'That change is reported to recover +0.314 judge accuracy on category-2 (temporal) questions in the matched set.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2608.11775, https://arxiv.org/abs/2608.11775. License: CC BY 4.0. v1 (12 Aug 2026). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

@misc{iil_note_mgeq17dy,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{The Sleeping Agent: What Gist-Based Context Compression Loses and Why}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2608-11775-sleeping-agent/},
  note = {Permanent ID iil:note:mgeq17dy. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2608-11775-sleeping-agent.md.