relore — GitHub project memory

Every codebase has lore. Ask it, then check it against the code.

relore indexes a repository's complete issue and pull-request history and keeps a working clone beside it, both served over one HTTP API. Three questions, in the order an agent hits them:

1. Is somebody already doing this? inflight <issue> — every open claimant, its state and its author, in one hop. The most expensive thing an agent does is patch something that is already in review.

2. Why is it like this? Why a fallback cannot be removed, which approach was tried and rejected, what a maintainer made the last five contributors change — none of it is in the codebase. thread, why PATH:LINE and search return it at comment level with who said it attached, so --trust authoritative keeps only what someone with write access settled and machine authors are excluded by default. search --symbol GemmaRotaryEmbedding asks what has been said about a function — a question grep cannot answer.

3. Is that still true? grep, symbol and copies run against that clone at HEAD, server-side, no checkout on your side. A three-year-old review is a claim about code that has moved since.

Threads tell you what people decided; the code verbs tell you whether it was true. A [contributor claim] confirmed by grep is stronger evidence than either alone — the argument for this over a search box.

Six agents have run it cold on the same transformers bug, one per release, on --help alone and with no memory of the runs before. They filed 39 issues against the tool; most are fixed, each run checking the last one's from the outside. #8 is the record.

Get started

• Deploy the service.
• Use relore.

What you get

The trade is freshness: each thread records its last indexing visit; later GitHub changes are unknown. Collector poll times do not guarantee freshness for every thread. Why not GitHub search has the measurements.

Running the daemon needs a Postgres and a GitHub token with issues:read + pull_requests:read — never write. deploy/ has a Helm chart and the scripts that are its interface; setup is in docs/operations.md.

Docs

• docs/how-search-works.md — what kind of index this is (lexical and ranked, no embeddings), and what it is better and worse at than grep.
• docs/cli.md — worked queries, and the empty results that are not faults.
• docs/why-not-github-search.md — the comparison.
• docs/architecture.md — what is indexed, what is deliberately not, the symbol lens, and language plugins.
• docs/agents.md — wiring it into a harness (not to be confused with AGENTS.md below, which is the contributor contract for this repository).
• docs/ranking.md — the scoring model and what the benchmark says.
• docs/operations.md — backfill, poll, clones, tokens, serving.
• docs/security.md — the properties and how each is enforced.
• docs/release.md — release branches, tags, and PyPI publishing.
• CHANGELOG.md — notable changes by release.
• docs/prior-art.md — related work, and what was borrowed.
• AGENTS.md — the operating contract: the invariants with tests behind them.

License

Apache-2.0.