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:
json{ "query": "what does the user like to eat" }
Parameters
| Parameter | Type | Required | Description |
|---|---|---|---|
query | string | Yes | Search query - keywords, phrases, or natural language questions |
type | string | No | Filter by memory type: raw, fact, summary, preference, context, or all. Default: fact |
subject | string | No | Filter to facts about specific person (e.g., "user", "Bob", "John") |
limit | number | No | Maximum results to return (default: 10, max: 50) |
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:
json{ "success": true, "output": "Found 3 memories:\n\n--- Result 1 (score: 0.850) ---\nContent: User enjoys Italian food...", "exitCode": 0 }
Response Fields
| Field | Type | Description |
|---|---|---|
success | boolean | True if search completed successfully |
output | string | Formatted search results or error message |
error | string | Error message if search failed |
exitCode | number | Process exit code (0 = success) |
Examples
Basic fact search
json{ "query": "user's favorite food" }
Search with subject filter
json{ "query": "birthday", "subject": "user" }
Search all memory types
json{ "query": "conversation about work", "type": "all", "limit": 20 }
Search for preferences
json{ "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
