remember-md/remember/ask

/remember:ask — Ask the Brain

Wraps the mcp__remember__search_brain MCP tool with: query reformulation, multi-call retry, citation-rich answer synthesis, and honest "not found" handling.

Prerequisite

The @remember-md/mcp server must be configured in the host's MCP config (added automatically by /remember:init since plugin v2.5.0). The tool mcp__remember__search_brain must be available in this session. If it isn't:

⚠️ The brain search MCP server isn't connected. Run `/remember:init` to
auto-configure, then restart Claude Code.

Don't proceed without the tool — there's no fallback in v1.

Step 1: Parse the user's question

The slash command is invoked as /remember:ask <question>. If the question is empty or only the slash command itself, ask the user what they want to know and stop.

Note the user's language (Romanian / English / mixed). Final answer will mirror the user's language unless content forces English (e.g., technical terms).

Step 2: Reformulate the query

Convert the natural-language question into a topical search phrase that matches how the brain is actually written. Apply these rules (also documented in the search_brain tool description — read it before this step):

• 3–7 words, declarative noun-phrase
• Strip stopwords + question words ("what", "how", "ce", "cum", "do I", "am I")
• Keep nouns, concepts, proper names
• Language match the content, not the question. Technical notes in this brain are usually English. Personal/strategic/relational notes are usually Romanian. If unsure, plan to try both languages (Step 3 retry).

Quick reformulation table

If the question is trivially answered by content already in the Persona block (Mission, Directives, recent Evidence Log), answer directly without calling the tool. Don't waste a tool call on Persona content.

Step 3: Call mcp__remember__search_brain

mcp__remember__search_brain({ query: "<reformulated phrase>", top_k: 8 })

Default top_k: 8. Use:

• 5–7 for focused queries (one specific fact/decision)
• 10–15 for broad surveys ("what do I know about X")
• 20+ for "give me everything on this topic"

Step 4: Evaluate result quality and retry if weak

Strong = top result has score >= 0.025 AND its path/heading_path/text obviously matches the user's intent.

Weak = top result score < 0.020, OR results are off-topic, OR note says vector index is building (BM25-only mode).

If weak, retry once or twice with alternative phrasing:

1. Try a different keyword set (synonyms, broader/narrower concepts)
2. Try the other language (RO ↔ EN)
3. For "decision" queries that miss, try the project name directly
4. For person queries that miss, try the person's project as second hit

Cap at 3 total search_brain calls per /remember:ask invocation. Don't loop.

Step 5: Synthesize the answer

Write a natural-language answer in the user's language (Romanian if they asked in Romanian, English if they asked in English; mix when the content is mixed).

Structure:

1. Direct answer in 1–2 sentences. Lead with the conclusion.
2. Supporting evidence — bullets or table with [[wikilinks]] citing the exact paths from results[].path. Quote concise snippets when they carry weight.
3. Caveats when relevant (data is stale, decision was for a specific project, contradiction noted, etc.)
4. Closing if useful — pointer to the source-of-truth file with strongest score.

Wikilink format: [[Notes/foo.md]] or [[Notes/foo.md|short label]]. Use the exact path from the tool result so Obsidian can resolve it.

Quote text field directly when concise (1–3 sentences). For longer chunks, paraphrase + cite. Never paraphrase technical details (paths, IDs, dates, names) — quote those verbatim.

Step 6: Honest "not found" handling

If after up to 3 search calls the brain genuinely doesn't contain a relevant answer:

Brain-ul nu pare să aibă informații despre <topic>.
 
Cele mai apropiate hits au scor sub 0.02 și sunt off-topic
(top: [[<path>]] — <heading_path>).
 
💡 Want me to remember this? Reply: `remember this: <whatever the user wanted to capture>`

Never invent content from outside the brain. If the brain is silent, say so. Adjacent or off-topic results should not be presented as if they answer the question.

Step 7: Confirm format

Final answer should always include:

• ✅ Direct answer (not just dump of results)
• ✅ Wikilink citations using exact result paths
• ✅ User's language (don't switch to English when they asked in Romanian)
• ❌ No hallucinated content beyond what the tool returned
• ❌ No raw JSON of tool output dumped at the user

Worked examples

Example A — entity lookup

User: /remember:ask cine este cezar

Reformulated: cezar

Tool call: mcp__remember__search_brain({ query: "cezar", top_k: 5 })

Answer (Romanian, since user asked in Romanian):

[[People/cezar.md|Cezar]] este [synthesize from result text]. Lucrează la [[Projects/X|X]]. Ultimul context relevant: [quote from result].

Surse: [[People/cezar.md]], [[Journal/2026-…|recent journal]].

Example B — decision recall

User: /remember:ask ce am decis despre netopia la 3dtoys

Reformulated: netopia 3dtoys exclusive payment provider

Tool call: above. If implementation plan ranks higher than decision note: Retry: decision netopia only 3dtoys

Answer:

Ai decis ca Netopia să fie singurul provider de plată pentru 3DToys ([[Notes/decision-netopia-only-3dtoys.md]]). Motivare: [quote].

Implementare în [[Projects/3dtoys/projects/3dtoys-site/NETOPIA_IMPLEMENTATION_PLAN.md]].

Example C — empty result

User: /remember:ask what do I think about react native mobile apps

Reformulated: react native mobile app preferences

Tool call: returns only 2 future-plan mentions (Remember GTM Phase 3), top score 0.0167.

Answer:

Brain-ul nu are o opinie capturată despre React Native mobile apps.

Singurele 2 mențiuni sunt planuri viitoare pentru un app mobil al Remember.md ([[Projects/remember/remember.md]] și [[Journal/2026-02-08.md]]) — nu o preferință personală.

💡 Want me to remember this? Reply: remember this: ...

Error handling

Anti-patterns

• ❌ Don't dump raw JSON of search_brain output at the user. Synthesize.
• ❌ Don't translate citations — wikilinks must use exact result paths.
• ❌ Don't loop beyond 3 search calls per ask invocation.
• ❌ Don't claim knowledge the tool didn't return. Brain-silence is honest output.
• ❌ Don't translate user's language in the answer just because the brain content is the other language — quote in original, summarize in user's language.
• ❌ Don't re-query for Persona content — it's already in your session context.