smart-memory-query_skill

Overview

This skill enforces proactive, query-optimized memory_search usage to avoid missed context and inefficient lookups. It triggers searches when prior context is referenced, a new task begins, or a proper noun appears. The skill builds short, focused queries by splitting intent into 2–4 token queries to reduce empty AND-based full-text search results. It merges and retries results with a fallback proper-noun query when needed.

How this skill works

On each user input the skill inspects for three triggers: references to prior decisions or history, the start of a new task/topic, and appearance of proper nouns (projects, tools, people). When any trigger fires, it decomposes the search intent into 2–3 independent angles, extracts 2–3 core tokens per angle, and runs a separate memory_search call for each query. If all queries return empty, it retries once using a single key proper noun to recover likely matches.

When to use it

• User references previous decisions, agreements, or history (e.g., “we decided this before”).
• A new task or topic is started and prior preferences or constraints may matter.
• A proper noun (project, tool, service, person) appears in the message.
• Unclear or ambiguous context where a missed memory is costlier than an extra search.
• When multiple triggers apply, run separate searches per trigger rather than one combined query.

Best practices

• Split intent into 2–3 independent angles instead of one large query.
• Limit each query to 2–4 tokens focused on nouns; prioritize proper nouns.
• Run one memory_search call per angle and merge results before acting.
• If all angle queries are empty, retry once with a single key proper noun.
• Prefer an extra targeted search over assuming no prior context exists.

Example use cases

• User asks about a past decision: run searches like “iCloud download setting” and “problem-solving preference.”
• Before starting a payment integration task: run “Paddle payment” and “payment integration decision.”
• User names a tool with an issue: run “OpenClaw search” and “search quality tuning.”
• Reference to migration plus model name: run “bge-m3 migration result” and “embedding model change.”
• New project setup: run “documentation structure preference” and “project template.”