Automated Research Literature Tracker

Academics, researchers, and technically curious people read a lot of papers. The problem isn’t finding relevant research — it’s staying current as it publishes. New papers on your topic appear constantly. Conferences release proceedings quarterly. Preprints surface on arXiv before formal peer review. You either read everything and burn out, or miss things and risk duplicating work someone already published.

OpenClaw can serve as your research monitoring layer. Define your interests once, and it watches the sources you care about, surfaces what’s relevant, captures it to your reference library, and keeps you current without you having to actively go looking.

What This Solves

1. Staying current — knowing when new papers matching your interests are published, before you stumble across them in a literature review three months later
2. Source fatigue — not having to manually check arXiv, PubMed, Semantic Scholar, and Google Scholar separately every week
3. Capture fatigue — downloading PDFs, renaming them consistently, adding metadata, and filing them in your reference manager by hand
4. Relevance filtering — distinguishing between a paper that mentions your keyword once in the introduction versus one that directly addresses your research question
5. Citation tracking — knowing when a paper you saved gets cited, and by whom

How It Works

Define Your Research Profile

You configure this once, with enough specificity to be useful:

~/research/tracker-config.yaml

interests:
  - name: "federated-learning-convergence"
    query: "federated learning convergence rate"
    sources: [arxiv, pubmed]
    categories: [cs.LG, stat.ML]
    keywords: [federated learning, distributed optimization, convergence]
    exclude: [wireless, communication efficiency]  # not my subfield
    min_relevance: high  # only alert on strong matches

  - name: "differential-privacy-ml"
    query: "differential privacy machine learning"
    sources: [arxiv, pubmed, semantic_scholar]
    keywords: [differential privacy, privacy-preserving, dp-ml]
    alert_threshold: medium  # alert on medium+ relevance

  - name: "ml-healthcare-fairness"
    query: "machine learning healthcare fairness bias"
    sources: [arxiv, pubmed]
    keywords: [healthcare ML, algorithmic fairness, bias detection]
    track_citations: true

arxiv:
  categories:
    - cs.LG
    - stat.ML
    - cs.AI
  update_schedule: "0 9 * * 1,4"  # Mon and Thu mornings

zotero:
  library_id: your-library-id
  api_key: your-zotero-api-key
  collection: "Research Tracker"

alert_schedule: "0 10 * * 1"  # Weekly Monday digest

What OpenClaw Does

Source monitoring — On your schedule, OpenClaw queries your configured sources:

arxiv: cs.LG, stat.ML — 47 new papers since last check
  → 12 match "federated-learning-convergence" (filtered: 9 excluded by keyword)
  → 3 match "differential-privacy-ml"
  → 1 matches "ml-healthcare-fairness" (high relevance → immediate alert)

pubmed: new entries matching "differential privacy machine learning" — 2 papers

Relevance scoring — OpenClaw reads titles, abstracts, and keyword matches to score relevance:

📄 FEDERATED LEARNING — HIGH RELEVANCE
"Adaptive Convergence Rates in Heterogeneous Federated Networks"
arXiv:2403.12345 | Authors: Liu, Chen, Wang
Relevance: ████████░░ 8/10

Abstract:
We study convergence bounds for FedAvg under non-IID data distributions,
proving adaptive rates that depend on local gradient diversity rather than
global variance. Our approach achieves 40% faster convergence vs baseline.

Key matches: federated learning ✓, convergence ✓, non-IID ✓
Notes: Uses similar setup to Wang et al. (2023) — check for novelty
PDF: captured to Zotero / Research Tracker / federated-learning-convergence/

Capture to Zotero — Papers above your threshold get captured automatically:

✅ Captured to Zotero:
  "Adaptive Convergence Rates in Heterogeneous Federated Networks"
  → Collection: Research Tracker / federated-learning-convergence
  → Tags: federated-learning, convergence, non-iid, arxiv-2024
  → Notes: Added flagged similarity to Wang et al. (2023)

Weekly digest — Every Monday, you get a structured summary:

📚 RESEARCH DIGEST — Week of Apr 21, 2026

FEDERATED LEARNING CONVERGENCE (3 new)
  ★★★ "Adaptive Convergence Rates in Heterogeneous Federated Networks"
      arXiv:2403.12345 — Liu, Chen, Wang — HIGH relevance
      Key finding: 40% faster convergence via gradient diversity bounds
      Action: Read before citing — may affect your convergence analysis
  
  ★★ "Compressed Federated Learning with Gradient Sparsification"
      arXiv:2403.11888 — Park et al. — MEDIUM relevance
      Communication-compression approach — tangential to your work
  
  ★ "FedNova with Partial Participation" — arXiv:2403.11901 — MEDIUM
      Minor contribution, likely not worth deep read

DIFFERENTIAL PRIVACY ML (2 new)
  ★★ "Privacy-Utility Tradeoffs in DP-SGD: A Large-Scale Empirical Study"
      arXiv:2403.12200 — Kim & Singh — MEDIUM relevance
      Large-scale evaluation of epsilon values — good background ref

ML HEALTHCARE FAIRNESS (1 new)
  ★★★ "Detecting Racial Bias in Clinical Risk Prediction Models"
      JAMA:2024.1847 — Roberts et al. — HIGH relevance
      Directly addresses your sub-question on bias detection methodology

CITATION ALERT
  "Gradient Diversity in Federated Systems" (saved Mar 2026) — cited by:
    → "Compressed Federated Learning" (above) — confirms your prior art search
  
⚠️ GAPS DETECTED
  No new papers on "personalized federated learning" in 3 weeks.
  Consider broadening query or checking if field has shifted focus.

Setting It Up

• OpenClaw with file access, web fetch, and a messaging channel
• arXiv API — free, no key required, rate-limited to 1 request per 3 seconds
• PubMed E-utilities — free, requires API key for higher rates (free tier available)
• Semantic Scholar API — free tier available, or use unofficial arXiv-to-SS links
• Zotero — API key and library ID from zotero.org/settings/keys
• Python/Scholarly (optional) — for more advanced academic search via CrossRef
• A weekly cron job — for the digest; daily checks are possible but generate more noise

What OpenClaw Can’t Do

It can’t read full papers and extract nuanced conclusions — for deep relevance, you still need to read the paper. It can help you decide which papers to read, not what they mean.

It can’t access paywalled papers directly unless you have institutional access configured. Preprints on arXiv and open-access papers on PubMed Central work well; Nature/Science papers behind paywalls require you to fetch them manually or configure proxy access.

It can’t replace a proper systematic review. For comprehensive literature surveys, you still need keyword-based searching and hand filtering. OpenClaw helps you stay current, not conduct exhaustive retrospectives.

Why This Works

Research moves faster than any individual can track. The problem isn’t finding papers — it’s knowing which ones matter and building the habit of checking. Most researchers rely on email alerts from journals, Google Scholar tracking, or periodic manual searches. All of these require the researcher to remember to check and do the filtering work themselves.

OpenClaw acts as the layer between you and the firehose — it does the checking, filters the noise, captures the relevant papers, and summarizes what’s new in a way you can process in five minutes on a Monday morning. The goal isn’t to read everything; it’s to never miss the paper that would have changed your direction.