← All notes

Summary

The context is reduced by 90-99% in token count, yet the model decides what to keep and what to drop with no guarantee. The operator has no fine-grained control over summary volume since prompt instructions are largely ignored. We introduce parallel compaction, a block-based design that gives the operator direct, fine-grained control over summary volume through block count. [S1]

Claims

• id: C1 claim: 'Sequential compaction is described as reducing context by 90-99% in token count.' source: S1 evidence: "the context is reduced by 90–99% in token count"
• id: C2 claim: 'They say the operator lacks fine-grained control over summary volume.' source: S1 evidence: "the operator has no fine-grained control over summary volume"
• id: C3 claim: 'They introduce a block-based parallel design that controls volume through block count.' source: S1 evidence: "a block-based design that gives the operator direct, fine-grained control over summary volume"

Verification

• Code available: No official repository URL for this paper was found on the abs page or in the HTML text searched on 2026-11-11.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live note ao-compaction-memory-and-parallel. Remory in this batch supplements a summary with learned tokens rather than changing block structure.
• Limits: No official code URL was found. The 90-99% figure describes typical sequential token-count reduction, not a measured quality score. End-to-end table cells are not restated here unless they appeared in the extract as prose.
• 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 22 May 2026 (v1). Sources: [S1] https://arxiv.org/abs/2605.23296, v1 (22 May 2026). checked_on: 2026-11-11.

When to apply

Use when asking how to control summary size during compaction for long-running agent serving, rather than leaving volume to a single model summary.

Facts

• id: F1 claim: 'The authors say sequential compaction can reduce context by 90-99% in token count, with no guarantee of what is kept.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'They say the operator has no fine-grained control over summary volume under sequential compaction because prompt instructions are largely ignored.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'Parallel compaction is a block-based design that sets summary volume through block count, compared with sequential compaction on HotpotQA and LoCoMo.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2605.23296, https://arxiv.org/abs/2605.23296. License: CC BY 4.0. v1 (22 May 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_pcz5a4cz,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{Parallel Context Compaction for Long-Horizon LLM Agent Serving}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2605-23296-parallel-context-compaction/},
  note = {Permanent ID iil:note:pcz5a4cz. 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/2605-23296-parallel-context-compaction.md.