mmq

name	mmq
description	Query the knowledge base. Reads session context first, then the root index, then relevant atlas, concept, and synthesis pages, and synthesizes a clear answer with [[wikilink]] citations. Use when you want to ask about past decisions, patterns, lessons, or technical knowledge captured in previous AI conversations. Use the `insights` argument to surface workflow-coaching advisories from hook-generated data.
argument-hint	[question | insights]
allowed-tools	Read Glob Grep

Objective

Answer a knowledge-base question using index-guided retrieval and cite sources with [[wikilinks]].

This command is operational only. Do not write /mmq, /memory-mason:mmq, or their execution chatter back into the vault.

Path Resolution

Before any other reasoning, resolve vault config in this priority order:

1. Project ./.env
2. Project ./memory-mason.json
3. Global ~/.memory-mason/.env
4. Global ~/.memory-mason/config.json

Resolve {vault} (absolute path) and {subfolder} from the matched source; .env sources use MEMORY_MASON_SUBFOLDER (default: ai-knowledge), JSON sources use subfolder. If none provide a vault path, state the knowledge base is not initialized.

Use these paths:

• Session context: {vault}/{subfolder}/_meta/context.md
• Manifest: {vault}/{subfolder}/_meta/manifest.json
• Build log: {vault}/{subfolder}/_meta/log.md
• Index: {vault}/{subfolder}/index.md
• Atlas: {vault}/{subfolder}/atlas/
• Atlas bases: {vault}/{subfolder}/atlas/bases/
• Concepts: {vault}/{subfolder}/concepts/
• Synthesis: {vault}/{subfolder}/synthesis/
• Sessions: {vault}/{subfolder}/sessions/
• Raw captures: {vault}/{subfolder}/_raw/

Steps

1. Read {vault}/{subfolder}/_meta/context.md if it exists.

• Treat context.md as a fast cache of recent context, not a replacement for the knowledge base.
• If context.md directly answers the question with enough precision, answer immediately.
• If context.md is relevant but incomplete, extract likely page names, tags, dates, and follow-up questions from it, then continue.

2. Read {vault}/{subfolder}/index.md.

• This is the primary guided retrieval mechanism after the session context.
• If it is missing, state that the knowledge base is not initialized.

3. Read the index table carefully.

• Identify 3-10 pages most relevant to the user question.
• Prefer pages explicitly mentioned in context.md when they are relevant.
• Use atlas pages for broad topic questions, synthesis pages for cross-cutting pattern questions, concept pages for specific definition, workflow, or decision questions, and session pages for temporal or episodic questions.

4. If index selection yields insufficient coverage, run a grep stage.

• Search concepts/, synthesis/, atlas/, and sessions/ for the query keywords (case-insensitive). Also check aliases: and tags: frontmatter fields for keyword matches.
• Add any newly identified relevant pages to the reading list. Prefer concept and synthesis pages over session pages for durable factual questions.

4a. For temporal questions ("what did I do last week", "what happened on YYYY-MM-DD", "recent activity"), read the sessions tier.

• Glob {vault}/{subfolder}/sessions/ for notes matching the relevant date range (filter by the date frontmatter field).
• Use atlas/bases/sessions-timeline.base as a structural guide to understand sort order, but read the actual session .md files for content.
• Cite session notes in the answer. Confidence aggregation does not apply to session notes.

5. Read selected pages in full.

• Pull complete article context before answering.
• If the answer depends on source provenance or freshness, read {vault}/{subfolder}/_meta/manifest.json to see which _raw/YYYY-MM-DD/ sources were compiled.
• If needed, inspect specific raw chunk files directly (both legacy _raw/YYYY-MM-DD/NNN.md and session-scoped _raw/YYYY-MM-DD/HHMMSS-{sid8}-NNN.md) instead of assuming the vault is current.

6. Aggregate confidence from cited pages.

• After reading concept pages, collect the confidence field from each cited concept's YAML frontmatter. Valid values: high, medium, low.
• Apply lowest-wins semantics: if any cited concept is low, overall confidence is low; else if any is medium, overall confidence is medium; else high.
• If any cited concept has status: seedling or contains a [!gap] callout, note it in the answer.
• Append to the answer footer: Confidence: [high|medium|low] (based on N cited concept pages) If any seedling or gap concepts were cited, add: Note: answer draws on N seedling/gap concept(s): [names].
• Apply only to cited concept pages. Do not aggregate from atlas, synthesis, or session pages.

7. Synthesize a clear, thorough answer.

• Cite supporting sources with [[wikilinks]] using the {subfolder}/ prefix: [[{subfolder}/concepts/example-concept]], [[{subfolder}/synthesis/example-tag]], [[{subfolder}/atlas/example-tag]], or [[{subfolder}/sessions/YYYY-MM-DD-{sid8}]].
• Use context.md only as retrieval assistance. Do not cite _meta/context.md as authoritative when a durable page exists.
• Mention raw chunk paths in prose only when you directly inspected them and no durable page exists yet.

8. Handle missing knowledge honestly.

• If the knowledge base does not contain relevant information, say so clearly.
• Suggest running /mmc if relevant information appears to exist in _raw/ but is not compiled into durable pages yet.
• Use {vault}/{subfolder}/_meta/log.md only to clarify recency or whether a compile happened. Do not treat the build log as a primary knowledge source.

Predefined arguments

Some arguments map to predefined retrieval flows instead of free-form questions. When the argument exactly matches one of these tokens (case-insensitive, whitespace-trimmed), follow the flow defined here. Otherwise treat the argument as a free-form question and run the default ## Steps flow.

insights

Surface current workflow-coaching advisories from the hook-generated knowledge.

Steps:

1. Read {vault}/{subfolder}/atlas/workflow-coaching.md. If the file is missing or empty, respond with: No workflow-coaching advisories yet. Hooks have not crossed the nag threshold. and stop.
2. Parse the ## Active Advisories section.
3. List up to the top 5 entries by count descending, then by most recent last descending.
4. For each surfaced advisory, render one bullet:
   • **<count>x** \` (kind: ) — first: , last: followed by a one-line note describing the workflow signal. If the advisory has asnippet, append it: — ""`. Omit the trailing snippet segment entirely when absent (older advisories predate this field).
5. Append a footer: See [[{subfolder}/atlas/workflow-coaching]] for the full list.
6. Do NOT write back to the vault. This is a read-only retrieval.

Candidate skills

After the advisory list, append a ### Candidate skills subsection. Populate it from the same ## Active Advisories data:

• Include every advisory whose kind is not prompt-repeat.
• Include every prompt-repeat advisory whose count is 5 or more.
• Phrase each entry as a recommendation rather than a raw signal, and branch the phrasing by kind:
  • prompt-repeat — a repeated prompt pattern suggests automation or a dedicated skill could eliminate the manual repetition.
  • error-repeat — a recurring error suggests the underlying bug or misconfiguration should be fixed, not turned into a skill.
• Render one bullet per candidate:
  • **<count>x** \` (kind: ) — (e.g. "Recurring task: consider creating a skill or automation for this workflow." forprompt-repeat; "Recurring error: investigate and fix the underlying cause." for error-repeat). If the advisory has a snippet, append it in quotes so the recommendation is actionable: — ""`. Omit when absent.
• If no advisories meet the inclusion criteria, omit the subsection entirely.
• Keep the existing advisory list format unchanged above this subsection.

Confidence aggregation does NOT apply to insights because advisories are not concept pages. Skip the Confidence: footer.