---
source: https://iinventedlight.com/library/notes/ao-compaction-summarize-and-resume.md
title: "Summarize and resume: long-horizon agents without bigger windows"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; CC0 1.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, [2509.13313] ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization, CC0 1.0; arXiv, [2510.00615] ACON: Optimizing Context Compression for Long-horizon LLM 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 rerun; only the fetched arXiv abstract pages were read. Results are as reported by the authors."
category: agent-operations
topic_area: "agent operations: compaction"
tags: agents, compaction, summarization, long-horizon
sources:
  - id: S1
    title: "[2509.13313] ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization"
    publisher: arXiv
    url: https://arxiv.org/abs/2509.13313
    accessed: 2026-11-04
    license: "CC0 1.0"
  - id: S2
    title: "[2510.00615] ACON: Optimizing Context Compression for Long-horizon LLM Agents"
    publisher: arXiv
    url: https://arxiv.org/abs/2510.00615
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# Summarize and resume: long-horizon agents without bigger windows

## Short answer
Periodically condense history into a summary and continue from it. ReSum does this without retraining and gained 4.5% over ReAct. ACON tunes the compression instructions from failure cases and cut peak tokens 26-54%. [S1][S2]

## Practice
Review what your summaries drop by looking at failed runs, and refine the summary prompt. [S2]

## Claims
- id: C1
  claim: ReSum periodically condenses interaction histories.
  source: S1
  evidence: "periodically invoking an external tool to condense interaction histories into compact summaries"
- id: C2
  claim: ReSum improved 4.5% over ReAct without training.
  source: S1
  evidence: "ReSum achieves a 4.5% improvement over ReAct in training-free settings"
- id: C3
  claim: ACON refines guidelines from failure analysis.
  source: S2
  evidence: "iteratively refines compression guidelines based on failure analysis of the agent"
- id: C4
  claim: ACON reduces peak tokens 26-54%.
  source: S2
  evidence: "reduces peak token usage by 26-54%"

## Sources
- [S1] [2509.13313] ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization. arXiv. https://arxiv.org/abs/2509.13313 . License: CC0 1.0. Accessed 2026-11-04.
- [S2] [2510.00615] ACON: Optimizing Context Compression for Long-horizon LLM Agents. arXiv. https://arxiv.org/abs/2510.00615 . 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).