Knowledge Persistence Architectures for AI-Based Coding Agents

Replication package for a systematic literature review following PRISMA 2020 on knowledge persistence architectures for AI-based coding agents. The package contains the protocol, search strategy, saved database exports, citation-chasing records, deduplication ledger, screening decisions, structured extraction data, synthesis tables, and PRISMA flow source.

Author: Alexandre Zatti
Affiliation: Programa de Pos-Graduacao em Computacao Aplicada (PPGPCA), Universidade Federal da Fronteira Sul (UFFS)
License: CC-BY-4.0
Version: 2.0.1 Zenodo: 10.5281/zenodo.20138710 Zenodo concept DOI: 10.5281/zenodo.19463131

Summary

The review used Scopus, arXiv, and Semantic Scholar citation chasing. The final arXiv cutoff was submittedDate:[202301010000 TO 202604272359], and final Scopus branches used PUBYEAR > 2022. The Scopus method is branch-split because the combined query returned HTTP 413 during validation.

Quantity	Count
arXiv raw rows	1,118
Scopus raw rows	130
Database raw rows	1,248
Citation-chasing unique records	1,071
Rows read by final deduplication	2,319
Duplicate rows removed or marked	484
Title/abstract screening records	1,835
Reports sought for retrieval	165
Retrieved full texts	163
Publisher-abstract-resolved records	1
Reports not retrieved and not assessed	1
Reports assessed for eligibility	164
Full-text included records	95
Full-text excluded records	69

Synthesis denominators:

Denominator	N	Use
Included studies	95	All full-text included records.
Architecture studies	85	Architecture taxonomy, mechanisms, quality, cost, adverse evidence, and synthesis.
Contextual/boundary studies	10	Contextual, benchmark, security, or boundary records reported separately.
Strict controlled-comparison studies	56	Same-system no-persistence or memory-ablation claims.
Broad controlled-comparison studies	76	Strict comparisons plus external or non-memory architecture baselines.

Repository Contents

protocol/       Protocol, PRISMA checklist, and amendments
search/         Search strategy, execution logs, scripts, and saved exports
screening/      Criteria plus title/abstract and full-text decisions
extraction/     Extraction schema, compiled TSV, per-record JSON, and digests
analysis/       Validation scripts and generated synthesis tables
synthesis/      Synthesis matrix and PRISMA flow summary
references/     BibTeX entries for all 95 included records

Full-text PDFs are not redistributed in this package. The package stores retrieval metadata, screening decisions, extraction data, factual digests, and citations instead.

Key Files

• Search strategy: search/search-strategy.md
• arXiv rows: search/arxiv-query-chunk-results.tsv
• Scopus branch strings: search/scopus-query-branches.tsv
• Citation chasing records: search/results-citations.json
• Deduplication ledger: search/dedup.tsv
• Title/abstract screening: screening/title-abstract-screening.tsv
• Full-text screening: screening/full-text-screening.tsv
• Included studies ledger: screening/included-studies.tsv
• Extraction TSV: extraction/extracted-data.tsv
• Extraction JSON files: extraction/json/
• Factual digests: extraction/digests/
• Synthesis matrix: synthesis/synthesis-matrix.md
• PRISMA flow summary: synthesis/prisma-flow.md

Reproducing Counts

The final PRISMA and synthesis counts should be reproduced from saved artifacts, not from live database calls. Live arXiv, Scopus, or Semantic Scholar queries may return different counts after the execution date.

Recommended checks:

python analysis/recompute_counts.py
python analysis/audit_row_level_synthesis.py
python analysis/check_bibliography_coverage.py

Expected results:

• recompute_counts.py: 95 included, 85 architecture, 10 contextual/boundary.
• audit_row_level_synthesis.py: 35 checks, 0 warnings, 0 failures.
• check_bibliography_coverage.py: 95 present, 0 missing.

Citation

The version-specific Zenodo DOI for this release is 10.5281/zenodo.20138710. The concept DOI 10.5281/zenodo.19463131 resolves to the latest Zenodo version.

Cite this release as:

@misc{zatti2026slrreplication,
  author    = {Zatti, Alexandre},
  title     = {{Knowledge Persistence Architectures for AI-Based Coding Agents: Replication Package}},
  year      = {2026},
  version   = {2.0.1},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20138710},
  url       = {https://doi.org/10.5281/zenodo.20138710}
}

License

This work is licensed under Creative Commons Attribution 4.0 International (CC-BY-4.0). Original publications referenced in references/included-studies.bib remain under their respective authors' copyrights and are not redistributed here.