vault-indexlisted

# Fill the vault index The MemPalace index is **summary cards**, one per note: name, path, what it answers, terms, links. It exists to find a note when the exact word is unknown: grep will never find "where was the thing about the agent getting dumber on long tasks", because the note does not contain "dumber". - Vault: `${user_config.vault_dir}` - Script: `${CLAUDE_PLUGIN_ROOT}/scripts/vault_index.py` (runs on any Python; it re-launches itself in the MemPalace tool environment, where chromadb and the embedder live) **Do not fill it with `mempalace mine`.** Measured on a real 1100-note vault: `mine` added 1043 drawers of raw text, including an inbox with a list of films, and eight test queries returned byte-identical results before and after those drawers were removed. The index is made by the summary, not by volume. **Do not depend on the `mempalace_*` MCP tools here.** The server starts with the app; when it did not connect, a task waiting for its tools hangs. The script writes directly, with the palace's own embedder (a different embedder silently breaks search for the whole palace). ## Steps **1. Take a batch.** ```bash uv run --no-project "${CLAUDE_PLUGIN_ROOT}/scripts/vault_index.py" gaps --vault "${user_config.vault_dir}" uv run --no-project "${CLAUDE_PLUGIN_ROOT}/scripts/vault_index.py" gaps --vault "${user_config.vault_dir}" --batch 40 ``` 30 to 50 notes per batch. More, and the quality of summaries drops faster than speed grows. Stubs (a title and empty hea