← All notes

Summary

Compaction can drop standing user constraints (side constraints, or SCs). The authors report that evaluated compactors retained only 17% of injected SCs on average, and that an inbox agent deleted emails without approval in 30% of sessions after compaction. An SC extractor that keeps constraints separate from the task summary and appends them after compaction reached over 90% retention in their controlled scenarios. [S1]

Claims

• id: C1 claim: 'Evaluated compactors retained only 17% of injected side constraints on average.' source: S1 evidence: "The evaluated compactors retain only 17% of injected SCs on average."
• id: C2 claim: 'A separate SC extractor reached over 90% retention in their controlled scenarios.' source: S1 evidence: "achieving over 90% retention across all three scenarios"
• id: C3 claim: 'The authors point to a public code and data repository.' source: S1 evidence: "Code and data are available at https://github.com/ZhiqiEliWang/compaction-integrity"

Verification

• Code available: Paper: code and data at https://github.com/ZhiqiEliWang/compaction-integrity (HTTP 200 on 2026-11-11). Not run.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live practice note ao-compaction-side-constraints covers the same paper at a practice level. This is the paper-verification note.
• Limits: Author-reported. The 30% inbox figure is for their experiment, not a universal rate. The extractor result is for three controlled scenarios.
• 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 31 Jul 2026 (v1); last revised 11 Oct 2026 (v2). Sources: [S1] https://arxiv.org/abs/2608.11242, v2 (11 Oct 2026). checked_on: 2026-11-11.

When to apply

Use when asking what happens to standing instructions after an agent compacts its context, or whether a separate constraint list can restore them.

Facts

• id: F1 claim: 'Evaluated compactors retained only 17% of injected side constraints on average.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'An SC extractor that stores constraints separately from the summary reached over 90% retention in the authors'' three controlled scenarios.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'After compaction, their inbox agent deleted emails without approval in 30% of sessions; restating the constraint in the first later user turn reduced that to 13%.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2608.11242, https://arxiv.org/abs/2608.11242. License: CC BY 4.0. v2 (11 Oct 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_1fsfqdke,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{Lost in Context Compaction: When LLM Systems Drop Side Constraints}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2608-11242-lost-in-context-compaction/},
  note = {Permanent ID iil:note:1fsfqdke. 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/2608-11242-lost-in-context-compaction.md.