where-did-i-put-it

name	where-did-i-put-it
description	Finds a misplaced item by searching the user's memory store for prior "I put X in Y" statements; if nothing matches, offers to remember the current location instead.

Where Did I Put It?

Trigger on "where did I put", "where's my", "where is my", "どこに置いた", "どこ行った", "どこだっけ". Paired with the long-standing user memory tools (remember / recall / search_memory / semantic_memory_search) — this skill wraps them in a single conversational flow so users don't need to remember tool names.

Default flow

1. Extract the search target from the utterance — the noun after "my" / "the" / "わたしの" / "私の". If no clear target, ask once: "What item are you looking for?" / "何を探してますか？" — single clarifying question, not a dialog tree.
2. Primary lookup: call semantic_memory_search { query: "<item> location" } with limit: 3. Semantic search wins here because users say "keys" but previously stored "car keys on the shelf"; keyword search would miss that.
3. Literal lookup: fall back to search_memory { query: "<item>" } if semantic returned zero results. Some users store notes like "umbrella -> hallway closet" which only substring search surfaces.
4. Reply shape:
   • Exactly one result → read it verbatim. Do not paraphrase. Keys on the hook is what they wrote; that is what we repeat.
   • Multiple plausible hits → pick the most recent, then mention the count: "On the shelf (2 other matches in your memory)".
   • Zero hits → single-sentence fallback: "I don't have a note about <item>. Want me to remember where you put it?" If the user answers "yes" + location, call remember { text: "<item> is in <location>" } and confirm.
5. Never invent a location. The temptation for an LLM here is real — if memory is empty, say so and offer to store. Hallucinated locations are worse than admitting ignorance.

What this skill does NOT do

• No visual / camera search. Photos of cluttered desks aren't a reliable locator and the camera tool is too heavy-weight for a "where are my keys" flow.
• No Bluetooth tracker integration (AirTag, Tile, etc.). Those have their own apps; this skill wraps text-based memory only.
• No cross-device search. Multi-room memory sync is deliberately out of scope (see docs/privacy.md — memory stays on this tablet).

Follow-ups

• If the user replies "I moved it to X" after a hit, append a new remember entry with the new location and an implicit timestamp. Do not overwrite the old entry — memory is append-only so the user's own trail of where an item used to be stays intact.
• If the user replies "forget that" after any step, call forget { query: "<item>" } and confirm.