Recall knowledge base entries by meaning, not just keywords — hybrid search (lexical ripgrep + local vector embeddings + see-link graph) over `.claude/knowledge/entries/`. Use when looking for prior knowledge, decisions, pitfalls, or context that may be worded differently from the query (e.g. a Japanese query vs English identifiers, or synonyms the entry does not literally contain). Falls back to ripgrep-only when the vector index or its dependencies are absent. On-demand only — it is NOT wired into the per-prompt hook (that stays ripgrep for instant, model-free injection).
Record tacit knowledge — quirks, pitfalls, dependencies, decisions, root causes — as tagged Markdown entries in `.claude/knowledge/entries/`. Use this skill whenever discoveries are made during work, when the user shares undocumented system behavior, or at plan completion to capture lessons learned. Also use when Claude Code makes a mistake pointed out by the user — record what happened, why it was wrong, and what to do next time.
Review and maintain the knowledge base — find stale entries, orphan entries without connections, missing links between related entries, and generate topic summaries. Supports the "internalization" phase of knowledge management by surfacing knowledge for periodic review and reflection.
Persist multi-step plans and task progress across Claude Code sessions. Use when starting work that may span multiple sessions, resuming incomplete plans, or updating task progress. Supports two modes: Git-tracked (shared via commits) and Issue-centric (issue tracker as primary source of truth, local scratchpad for sessions).