README.md

metadata

license: cc-by-4.0
pretty_name: AgentArtifactCorpus (AAC)
language:
  - en
tags:
  - ai-agents
  - agent-memory
  - prompts
  - llm
  - knowledge-management
task_categories:
  - text-generation
  - text-classification
size_categories:
  - 100K<n<1M
extra_gated_prompt: >-
  AgentArtifactCorpus is released for research use under a Data Use Agreement.
  By requesting access you agree to: (1) use the data only for research; (2) not
  attempt to re-identify authors of the source repositories; (3) honour removal
  requests propagated from the opt-out registry; (4) preserve upstream file
  licenses and provide attribution under CC-BY-4.0; (5) cite the accompanying
  paper.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
  I agree to the Data Use Agreement: checkbox
configs:
  - config_name: corpus
    data_files: corpus/*.parquet
    default: true
  - config_name: manifest
    data_files: manifest/*.parquet

AgentArtifactCorpus (AAC)

Agent configuration artefacts (AGENTS.md, CLAUDE.md, .cursorrules, .mdc rules, system-prompt files) crawled from public GitHub repositories under permissive licences. AAC underpins the compaction, decomposition, and classifier experiments in the CIKM 2026 paper The Compaction Cliff in Long-Running AI Agent Memory.

• Code and reproduction: https://github.com/searchsim-org/knowledge-triage
• Datasheet: see DATASHEET.md in the code repository
• Access: gated, with a one-page Data Use Agreement (see above)

Configurations

Config	Contents	Use
corpus (default)	one row per artefact, with redacted text	training / analysis
manifest	same rows without text (hash + metadata)	full index, dedup and opt-out bookkeeping

from datasets import load_dataset
ds = load_dataset("searchsim/AgentArtifactCorpus", "corpus", split="train")

Schema

Rows are at the occurrence level: a template copied across many repositories appears once per repository, each with its own provenance. dup_count and is_duplicate let you deduplicate by content when you need to.

Column	Type	Notes
id	string	SHA-256 of the artefact content; the dedup key
repo_slug	string	anonymised pseudonym (anon/<hash>)
anonymized	bool	always True: every source repository is anonymised
file_path	string	path of the artefact within its repository
family	string	claude.md, agents.md, cursorrules, copilot, system-prompt, other
platform	string	crawler platform: claude, cursor, copilot, windsurf, continue, aider, universal
function	string	instruction, strategy, configuration, memory
authorship	string	human, template, agent, hybrid, unknown
specificity	string	e.g. project, global
size_bytes	int	original file size
char_count	int	original character count (from the crawl)
crawl_timestamp	string	ISO 8601 crawl time
dup_count	int	number of occurrences sharing this id
is_duplicate	bool	False on the first occurrence of an id, True on the rest
char_len	int	length of the redacted text (corpus config only)
redactions	int	number of PII/secret redactions applied (corpus config only)
text	string	redacted artefact text (corpus config only)

The manifest config carries every in-scope artefact (metadata only, no text), including artefacts whose body is not materialised. The corpus config carries the materialised, scrubbed subset with text.

Anonymity, provenance, and licensing

Source repositories are anonymised: every repo_slug is a stable pseudonym (anon/<hash>). The repo_license column records the upstream SPDX licence (not identifying). The dataset is distributed under CC-BY-4.0; cite the paper below rather than individual repositories.

Scrubbing

Every artefact passes these filters (see the code repository's build_full.py):

1. Secret scan — gitleaks / trufflehog / bearer-token shapes; any hit drops the file.
2. PII redaction — emails, GitHub @handles, internal hostnames, IPs, user paths, and attribution front-matter lines are replaced with <REDACTED_*> placeholders.
3. Anonymisation — the repo_slug is replaced with a pseudonym.

Opt-out

To have your repository removed, open a data-removal request on the code repository. Because slugs are anonymised, maintainers resolve the request through a private map held off-Hub, then publish a new gated revision (and Zenodo version) with the matching rows removed.

Citation

@inproceedings{zerhoudi2026compaction,
  author    = {Zerhoudi, Saber and Mitrovi\'{c}, Jelena and Granitzer, Michael},
  title     = {The Compaction Cliff in Long-Running AI Agent Memory},
  year      = {2026},
  booktitle = {Proceedings of the 35th ACM International Conference on Information and Knowledge Management},
  series    = {CIKM '26},
  doi       = {10.1145/3799682.3840567},
}