Project Memory

Overview

Answers a free-text question by interrogating everything the project already knows — extracted learnings (patterns/pitfalls/observations, active + archived), the architecture map (components and how they connect), the backlog (features and their status), and prior thinking output. It reads across all of these, ranks by relevance, and returns a synthesized answer with sources — instead of the user grepping four stores by hand.

This is the read-only, cross-store counterpart to shared/LEARNINGS-LOAD.md (which loads a fixed scoped slice into a build skill's context). Here the query is the user's question, and the whole memory is in scope.

Read-only: this skill never writes to .project/. It answers questions; it does not record them.

PHASE 0 — Pre-flight

1. Question: take it from the invocation argument. If empty, ask the user what they want to know (one short prompt) before continuing.
2. Project check: resolve $REPO to the main worktree (per shared/SYNC.md Worktree-aware Path Resolution) — the archive and stores live there, not in a feature worktree. If $REPO/.project/ does not exist, stop with: "No project memory here yet — run /core-setup first."
3. Language: answer in the runtime language from CLAUDE.md § User Preferences → Language: (see shared/LANGUAGE.md). This SKILL.md stays English; the user-facing answer follows their setting.

PHASE 1 — Gather (in parallel where possible)

Pull candidate evidence from each store. Keep each pull small — this is retrieval, not a dump.

1a) Learnings (active + archive) — the primary source. Run:

node ~/.claude/scripts/learnings-search.js "$REPO" search --query "<question>" --archive --json --cap 12

Returns scored entries {date, feature, type, source, author, summary, tags, archived, score}. These already blend active and archived memory, ranked by relevance (see shared/LEARNINGS-LOAD.md § Relevance model). If the question names a specific area, also pass --tags <tag> (a tag from LEARNING-WRITE.md § Tag Vocabulary) to sharpen the match.

1b) Architecture — read $REPO/.project/project-context.json → architecture.components[]. Keep components whose name, description, feature, endpoints[], or entities[] match the question tokens. For each kept component report name, layer, description, and its connects_to[] edges (so "how does X reach Y" is answerable). Also scan architecture.routes[] when the question is about routing/pages.

1c) Backlog — read $REPO/.project/backlog.json → features[] (fields via shared/BACKLOG-LOAD.md if the file is large). Keep features whose name/description/type match the question; report name, status, phase, and dependencies[]. This answers "is X built / planned / blocked?". Include archived features (.project/archive/backlog-archive.json) only if the question implies shipped work.

1d) Thinking output — Grep $REPO/.project/thinking/*.md for question tokens in filenames and H1 headings (per shared/DASHBOARD-CONTEXT.md § thinking-output — filenames are {date}-{type}-{slug}.md). List matching docs by title + date; do not inline their full body.

If a store is empty or absent, skip it silently — never fabricate.

PHASE 2 — Synthesize

Write a direct answer to the question, grounded only in what PHASE 1 returned. Structure:

1. Answer first — 1–3 sentences that actually answer the question. Lead with the conclusion.
2. Supporting detail — the relevant learnings/components/features in prose, not a raw dump. Prefer the highest-scored evidence; mention tags only when they aid the reader.
3. Sources: block — one line per claim-bearing item, so the user can verify:
   • learnings: learning [date] feature — summary (archived?)
   • architecture: component <name> (project-context.json#architecture)
   • backlog: feature <name> — <status> (backlog.json)
   • thinking: <title> (.project/thinking/<file>.md)
4. Gaps — if the memory has nothing relevant, say so plainly ("Project memory has nothing on X"); suggest where it would come from (a /dev-ship run, a /project-research doc). Do not pad with generic advice.

If a matching thinking doc looks central to the answer, offer to open it (the file is the source of truth, not a summary).

Notes

• Relevance, not recency: ranking comes from learnings-search.js scoreEntry() — tag > feature > keyword, recency only as a tiebreak. An old but on-topic entry will surface; that is the point.
• No writes, no side effects: if the user's question reveals a missing learning, note it in the answer — recording it is a separate /dev-ship concern, not this skill's.
• Token discipline: cap each store's pull (learnings --cap 12; a handful of components/features/docs). The answer is a synthesis, not a transcript.