tashfeenahmed/scallopbot/memory-search

Memory Search Skill

Search through stored memories using hybrid search combining BM25 keyword matching and semantic similarity. Returns ranked results with relevance scores.

When to Use

Use the memory_search skill for:

• Finding user facts: Recall what you've learned about the user
• Searching preferences: Find stored user preferences
• Retrieving context: Get relevant past conversation context
• Subject-specific queries: Find facts about specific people (user, friends, family)

Input Format

The skill accepts JSON arguments via the SKILL_ARGS environment variable:

{
  "query": "what does the user like to eat"
}

Parameters

Memory Types

• fact (default): Extracted facts about the user or others
• preference: User preferences and settings
• raw: Unprocessed conversation data
• summary: Condensed summaries
• context: Processed contextual information
• all: Search all memory types

Output Format

The skill returns JSON to stdout:

{
  "success": true,
  "output": "Found 3 memories:\n\n--- Result 1 (score: 0.850) ---\nContent: User enjoys Italian food...",
  "exitCode": 0
}

Response Fields

Examples

Basic fact search

{ "query": "user's favorite food" }

Search with subject filter

{ "query": "birthday", "subject": "user" }

Search all memory types

{ "query": "conversation about work", "type": "all", "limit": 20 }

Search for preferences

{ "query": "notification settings", "type": "preference" }

Search Behavior

• Hybrid search: Combines BM25 keyword matching with semantic similarity
• Default type: Searches "fact" type by default (not raw logs)
• Natural activation: Fresh memories surface easily; old memories fade but can return from a genuinely relevant topic/name without special history wording
• Real reinforcement: User confirmation strengthens memory; automatic retrieval does not
• User boost: Facts about the user get 1.5x score boost
• Relevance scoring: Results sorted by combined relevance score