memo-bank-query

Querying a memo-bank

A memo-bank is a read-only corpus of project docs served over MCP. It answers "what rules govern this code?" precisely, so you load the contract before you touch the code rather than after you've broken it.

In these projects specs are contracts and code follows them. If a spec governs the path you're editing, satisfy it; if you're about to violate it, stop and say so.

Is a memo-bank available here?

Look for .island-slices.json at the repo root, or mcp__memo-bank__docs_* tools. If neither exists, this skill doesn't apply — say so rather than inventing rules.

Two ways to reach it:

• MCP tools — mcp__memo-bank__docs_* (preferred when present).
• The installed CLI — memobank serve|validate|drift|coverage (often at .venv/bin/memobank); for one-off queries the Python API works too: memo_bank.load_federation(memo_bank.parse_registry(Path(".island-slices.json"))) then fed_resolve_path(fed, path) / fed_search_live(fed, query).

Pick the entry point by what you have

Editing a specific file → docs_resolve_path (the common case, ~2 reads)

1. docs_resolve_path(<repo-relative path>) → governing doc id(s), ordered by glob specificity (closest applies_to wins).
2. docs_get(<top id>) for the full contract, or docs_get_section(<id>, <section>) for one part (e.g. the restrictions).
3. Empty result = a coverage gap — that path is undocumented, not unconstrained. Say so plainly; don't fabricate a rule. The project's coverage loop will log it as spec-wanted.

A concept or question → expand first, then docs_search_live

Scoring is lexical bag-of-words: order and duplicates don't matter, and there is no substring match (auth does NOT match authentication). So the query must contain the corpus's actual words.

1. Expand with domain synonyms before searching. Add the concrete terms the docs plausibly use — e.g. for "login" → authentication auth login session token signature; for "crawling reviews" → refresh fetch ingest cache stale quota. This is the single biggest lever on hit quality (it has moved a top hit from a score of 3 to 32 in practice).
2. docs_search_live(<expanded query>) → ranked pointers; then docs_get the top hits.
3. Unsure of the vocabulary? docs_list first and reuse its tags.