← All notes

Summary

Remory adds a learned residual memory: after a text summary, a network appends a bounded sequence of soft tokens meant to help a frozen model approximate the continuation it would have produced from the full history. On SummHay, the authors say this improves source attribution at nearly unchanged insight coverage and approaches the full-context joint score using only 5.2% of the input positions. [S1]

Claims

• id: C1 claim: 'Remory is a neural memory network for context compaction.' source: S1 evidence: "We introduce Remory , a neural memory network for context compaction."
• id: C2 claim: 'On SummHay they report approaching the full-context score at 5.2% of input positions.' source: S1 evidence: "approaches the full-context joint score using only 5.2% of the input positions"
• id: C3 claim: 'They report consistent gains with residual memory on long-horizon agent benchmarks.' source: S1 evidence: "show consistent gains with residual memory"

Verification

• Code available: No official Remory repository URL was found on the abs page or in the HTML text searched on 2026-11-11. Cited third-party repos were not treated as this paper's code.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live note ao-compaction-memory-and-parallel. Parallel Context Compaction in this batch changes how summaries are produced, not residual tokens.
• Limits: No official Remory code URL was found on the abs page or in the HTML text. Agent-benchmark gains for Qwen3.8-27B and GLM-5.3-Flash are described as consistent; details sit in figures.
• 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 8 Oct 2026 (v1); last revised 21 Oct 2026 (v3). Sources: [S1] https://arxiv.org/abs/2610.11287, v3 (21 Oct 2026). checked_on: 2026-11-11.

When to apply

Use when asking whether a short learned memory can supplement a readable summary after compaction, instead of relying on the summary alone.

Facts

• id: F1 claim: 'Remory generates a bounded sequence of soft memory tokens, conditioned on the summary and appended after it.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'On SummHay, the authors say Remory approaches the full-context joint score using only 5.2% of the input positions.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'They report consistent gains with residual memory for Qwen3.8-27B and GLM-5.3-Flash on long-horizon agent benchmarks.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2610.11287, https://arxiv.org/abs/2610.11287. License: CC BY 4.0. v3 (21 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_q07n3cee,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{REMORY: Learning Residual Memory for Context Compaction}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2610-11287-remory/},
  note = {Permanent ID iil:note:q07n3cee. 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/2610-11287-remory.md.