---
source: https://iinventedlight.com/library/notes/ao-compaction-memory-and-parallel.md
title: "Beyond one summary: residual memory and parallel compaction"
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, [2610.11287] REMORY: Learning Residual Memory for Context Compaction, CC BY 4.0; arXiv, [2605.23296] Parallel Context Compaction for Long-Horizon LLM Agent Serving, 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, memory, latency
sources:
  - id: S1
    title: "[2610.11287] REMORY: Learning Residual Memory for Context Compaction"
    publisher: arXiv
    url: https://arxiv.org/abs/2610.11287
    accessed: 2026-11-04
    license: "CC BY 4.0"
  - id: S2
    title: "[2605.23296] Parallel Context Compaction for Long-Horizon LLM Agent Serving"
    publisher: arXiv
    url: https://arxiv.org/abs/2605.23296
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# Beyond one summary: residual memory and parallel compaction

## Short answer
A plain text summary loses detail and can stall the agent. Two fixes: add learned memory tokens after the summary (REMORY), or compact blocks in parallel for predictable size and less wait. [S1][S2]

## Reported effects
REMORY neared full-context scores using 5.2% of input positions and cut repeated tool calls. [S1] Parallel compaction reduced end-to-end wall time. [S2]

## Claims
- id: C1
  claim: A summary alone may not support every decision.
  source: S1
  evidence: "a textual summary alone may not support every subsequent decision"
- id: C2
  claim: REMORY uses 5.2% of input positions.
  source: S1
  evidence: "using only 5.2% of the input positions"
- id: C3
  claim: Blocking summarization stalls inference.
  source: S2
  evidence: "the blocking call stalls agent inference for tens of seconds"
- id: C4
  claim: Parallel compaction reduces wall time.
  source: S2
  evidence: "it reduces end-to-end wall time"

## Sources
- [S1] [2610.11287] REMORY: Learning Residual Memory for Context Compaction. arXiv. https://arxiv.org/abs/2610.11287 . License: CC BY 4.0. Accessed 2026-11-04.
- [S2] [2605.23296] Parallel Context Compaction for Long-Horizon LLM Agent Serving. arXiv. https://arxiv.org/abs/2605.23296 . 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).