---
title: 'DyadMem: A Long-Term Memory Benchmark of How Agents Work with Users'
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is CC BY 4.0 (verified on arXiv abs page 2026-11-11; icon by-4.0.png, http://creativecommons.org/licenses/by/4.0/); note is our own summary'
source: https://iinventedlight.com/library/notes/2610-03020-dyadmem.md
source_license: 'CC BY 4.0'
attribution: 'Summary by iinventedlight (AI-assisted). Source: Yifei Tao, Xinyu Zhong, Henry Hengyuan Zhao, Fanyi Wang, Tengda Guo, Wentao Qiu, Ying Wang, Liujian Tang, DyadMem: A Long-Term Memory Benchmark of How Agents Work with Users (CC BY 4.0).'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-11
status: published
approved_by: "Cole Withington, 2026-11-11 20:07 PT, Librarian chat widget t1715u: \"okay you can publish the ones you're comfortable with\""
checked_on: 2026-11-11
as_of: 2026-11-11
confidence: medium
untested: 'Not reproduced. Abs page and HTML key sections read 2026-11-11. Named code URLs checked for HTTP 200; code not run.'
category: arxiv
topic_area: 'ai-research: agent memory and context compaction'
tags: agents, memory, compaction, arxiv
arxiv_id: 2610.03020
alias: 2610.03020-dyadmem
sources:
  - id: S1
    title: 'DyadMem: A Long-Term Memory Benchmark of How Agents Work with Users'
    publisher: arXiv
    authors: 'Yifei Tao, Xinyu Zhong, Henry Hengyuan Zhao, Fanyi Wang, Tengda Guo, Wentao Qiu, Ying Wang, Liujian Tang'
    url: https://arxiv.org/abs/2610.03020
    version_read: 'v1 (2 Oct 2026)'
    accessed: 2026-11-11
    license: 'CC BY 4.0'
---
# DyadMem: A Long-Term Memory Benchmark of How Agents Work with Users

## Summary
Gold Memory QA ranges from 78.7% to 96.3%, yet Full Pipeline QA averages only 47.7% for open-weight models and 57.4% for proprietary models. DyadMem separately evaluates Capture, gold-conditioned standalone Update, Recall, Gold Memory QA, and Full Pipeline QA, making failures localizable rather than visible only through final-answer accuracy. Capture is precision-heavy and recall-poor: open-weight models average 54.0% precision but only 15.1% recall. [S1]

## Claims
- id: C1
  claim: 'Gold-Memory QA is reported between 78.7% and 96.3%.'
  source: S1
  evidence: "Gold-Memory QA ranges from 78.7% to 96.3%"
- id: C2
  claim: 'Full-Pipeline QA averages 47.7% for open-weight models and 57.4% for proprietary models.'
  source: S1
  evidence: "Full-Pipeline QA averages only 47.7% for open-weight models and 57.4% for proprietary models"
- id: C3
  claim: 'The HTML names a public GitHub repository.'
  source: S1
  evidence: "https://github.com/RenaissanceT/DyadMem"

## Verification
- Code available: HTML links https://github.com/RenaissanceT/DyadMem (HTTP 200 on 2026-11-11) and https://huggingface.co/datasets/RenaissanceT/DyadMem (HTTP 200). Not run.
- Reproduced: not reproduced (2026-11-11). Code not run.
- Related work: Live note ao-memory-long-term-benchmarks. AMA-Bench in this batch targets agentic application memory rather than user-dyad work.
- Limits: Author-reported stage averages. Capture is described as precision-heavy and recall-poor (open-weight 54.0% precision, 15.1% recall).
- 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 2 Oct 2026 (v1). Sources: [S1] https://arxiv.org/abs/2610.03020, v1 (2 Oct 2026). checked_on: 2026-11-11.

## When to apply
Use when asking where an agent memory system fails: writing memories, updating them, retrieving them, or answering once the right memories are present.

## Facts
- id: F1
  claim: 'DyadMem separately evaluates Capture, gold-conditioned Update, Recall, Gold-Memory QA, and Full-Pipeline QA.'
  confidence: medium
  date: 2026-11-11
  sources: [S1]
  license: CC BY 4.0
- id: F2
  claim: 'Gold-Memory QA is reported from 78.7% to 96.3%, while Full-Pipeline QA averages 47.7% for open-weight models and 57.4% for proprietary models.'
  confidence: medium
  date: 2026-11-11
  sources: [S1]
  license: CC BY 4.0
- id: F3
  claim: 'Capture for open-weight models is reported as 54.0% precision and 15.1% recall.'
  confidence: medium
  date: 2026-11-11
  sources: [S1]
  license: CC BY 4.0

## Sources
- [S1] arXiv 2610.03020, https://arxiv.org/abs/2610.03020. License: CC BY 4.0. v1 (2 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.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).