---
source: https://iinventedlight.com/library/notes/ao-compaction-side-constraints.md
title: "Context compaction can silently drop user constraints"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, [2608.11242] Lost in Context Compaction: When LLM Systems Drop Side Constraints, CC BY 4.0; arXiv, [2610.02163] AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents, CC BY 4.0."
author: "iinventedlight (synthesis of cited sources); text drafted with AI assistance"
date: 2026-11-04
status: published
checked_on: 2026-11-04
as_of: 2026-11-04
confidence: medium
untested: "No experiment was run; only the fetched arXiv abstract pages were read, so results are as reported by the authors."
category: agent-operations
topic_area: "agent-ops: context compaction"
tags: agents, context-compaction, constraints
sources:
  - id: S1
    title: "[2608.11242] Lost in Context Compaction: When LLM Systems Drop Side Constraints"
    publisher: arXiv
    url: https://arxiv.org/abs/2608.11242
    accessed: 2026-11-04
    license: "CC BY 4.0"
  - id: S2
    title: "[2610.02163] AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents"
    publisher: arXiv
    url: https://arxiv.org/abs/2610.02163
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# Context compaction can silently drop user constraints

## Short answer
When an agent compacts its context, standing user instructions ("do not delete anything until I confirm") can be lost. One study found compactors kept only 17% of injected constraints on average, and an inbox agent deleted emails without approval in 30 of 100 sessions after compaction. Keeping constraints in a separate list and re-appending them after each compaction restored over 90% retention. [S1]

## Practice
- Store standing constraints outside the summary and re-insert them verbatim. [S1]
- Decide deliberately when to compact and what working state to keep; learned compaction policies improved coding-agent pass rates. [S2]

## Claims
- id: C1
  claim: 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: Keeping constraints separate from the summary reached over 90% retention.
  source: S1
  evidence: "maintains constraints separately from the task summary and appends them to the compacted context"
- id: C3
  claim: Agents must decide when to compact and what state to preserve.
  source: S2
  evidence: "an agent must decide when to compact, what working state to preserve"
- id: C4
  claim: A judge reviewed compaction decisions to build training data.
  source: S2
  evidence: "use a judge to review its compaction decisions, summaries, and actions after compaction"

## Sources
- [S1] [2608.11242] Lost in Context Compaction: When LLM Systems Drop Side Constraints. arXiv. https://arxiv.org/abs/2608.11242 . License: CC BY 4.0. Accessed 2026-11-04.
- [S2] [2610.02163] AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents. arXiv. https://arxiv.org/abs/2610.02163 . License: CC BY 4.0. Accessed 2026-11-04.

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