| name | mneme-search |
| description | Use when the user asks a factual question whose answer might live in the vault. Invokes mneme_search. Production `mneme_search` is FTS5 BM25. The experimental feature-hashed lexical-vector backend is not wired into MCP. KG enrichment is gated to summarize or timeline when full-profile graph state is active. |
mneme-search
When the user asks a question that sounds like recall from prior work
or notes, search the vault first instead of guessing.
When to invoke
- "Did we decide X?"
- "What was the conclusion about Y?"
- "Show me everything I have on Z."
- The user invokes this skill directly.
How to invoke
Call the mneme_search MCP tool with:
query: the user's question in natural language.top_k: 5 by default. Raise to 10 only when the user asks for a
broad sweep.- Optional
filters.date_from / filters.date_to when the user
scopes their question to a specific time window.
Inspect the returned hits. Each has path, title, snippet, and
a relevance score. Read the snippets first. If they answer the
question, cite the path and reply. If they only partially answer,
follow up with mneme_recall on the most promising paths to pull the
full body.
What not to do
- Do not invoke this for the user's first hello or for questions whose
answer is general knowledge.
- Do not search for the same query twice in a row. If the first pass
returned nothing useful, refine the query terms or change the filter
window.
- Do not present raw snippets in your reply unless the user explicitly
asks for them. Summarize and link.
