dacorvo/transformers-gh-memory

huggingface/transformers issues and pull requests, as a funes memory

Every issue and pull request of huggingface/transformers with activity since 2018-01-01 — opening bodies, comments, reviews, inline review comments and PR diffs — chunked, embedded and written to a Lance table by funes, so the tracker can be searched by meaning and read back thread by thread. Kept fresh every few minutes by the funes-github Space.

Use it

Set funes up for your agent the usual way, once — funes add claude or funes add codex — and keep your own sessions as its default memory. Then tell the agent this memory exists, either way:

• —A skill, loaded when the topic comes up:

  hf download dacorvo/transformers-gh-memory --repo-type dataset --include 'skills/*' --local-dir ~/.claude   # Claude Code
  hf download dacorvo/transformers-gh-memory --repo-type dataset --include 'skills/*' --local-dir ~/.codex    # Codex

• —A paragraph in `CLAUDE.md` or `AGENTS.md`:

dacorvo/transformers-gh-memory is a funes memory of huggingface/transformers issues and PRs. For questions about transformers issues, PRs, reviews or who said what, pass it as the memory argument of the funes tools (recall, get, sessions). One session is one thread, gh/huggingface/transformers#<number>.

Then ask in plain words:

• —why are the experts sharded on the TP mesh for masked expert parallelism?
• —what was decided about byte fallback when loading the SentencePiece backend tokenizer?
• —read me the review of PR #48857

(funes itself: curl -fsSL https://huggingface.co/buckets/huggingface/funes/resolve/install.sh | sh)

What is in a thread

One thread is one funes session, gh/huggingface/transformers#<number>. Its turns, in order:

role is GitHub's relationship to the project — member, collaborator, contributor, first_time_contributor, none — or bot for automation (github-actions, the Copilot reviewer, the docs builder). The login stays in the text, so "what did ArthurZucker say about …" works. harness is github; the repo facet is huggingface/transformers.

What to expect

• —Append-only. funes never rewrites. An edited body is a new turn stamped with its edit time; the earlier text stays and recency ranks the newer one. Deleted comments stay. A lone comment edit does not move a thread's updated_at, so it lands at the thread's next activity.
• —Freshness. A thread updated on GitHub is here within about five minutes of its next activity.
• —Diffs are indexed in funes's deferred tier: recall over the discussion is unaffected; they answer "which PR touched this file".
• —Secrets. Text is redacted at index time and a fail-closed scanner holds back anything that slips through before it is published.

Provenance

Built by `gh_to_funes.py` from the GitHub REST and GraphQL APIs, into funes's turns-file format, then funes index and funes push. Embeddings: BAAI/bge-small-en-v1.5. The texts are their authors' words as published on GitHub, redistributed here for search; the diffs are code from huggingface/transformers (Apache-2.0). Fork the Space to build the same memory for another project.

<!-- funes:stats --> | | | |---|---| | Chunks | 1,164,131 | | Embedding model | BAAI/bge-small-en-v1.5 | | Updated | 2026-10-06 | <!-- /funes:stats -->