feat: Add semantic memory search endpoint with vector similarity

[elmorem]

Summary

• Add vector similarity search endpoint to Memory Service for semantic memory retrieval
• Implement pgvector-based cosine similarity search in repository layer
• Add search method to service layer with scope, topic, and confidence filtering
• Create SearchMemoriesRequest and SearchMemoriesResponse schemas
• Add POST /api/v1/memories/search endpoint with similarity scoring

Implementation Details

Repository Layer (memory_repository.py:344-395)

• Added search_by_similarity method using pgvector's cosine distance operator
• Converts distance (0-2 range) to similarity score (0-1 range) for intuitive results
• Filters out memories without embeddings
• Supports optional scope, topic, and min_confidence filtering

Service Layer (memory_service.py:338-366)

• Added search_memories method that wraps repository search
• Enforces include_deleted=False for production queries

API Layer (memories.py:128-160)

• POST /api/v1/memories/search endpoint
• Request includes query_embedding, optional scope/topic filters, limit (1-100), min_confidence
• Returns results with memory data and similarity scores
• Supports pagination with configurable result limit

Schemas

• SearchMemoriesRequest: Query embedding, optional filters, limit, min_confidence
• MemorySearchResult: Individual result with memory data and similarity score
• SearchMemoriesResponse: List of results with total count and limit

API Gateway

The existing API Gateway catch-all route (/api/v1/memories/{path:path}) automatically proxies requests to the new search endpoint with no additional configuration needed.

Test Plan

• Type checks pass (mypy) for memory service
• Code formatting applied (black)
• Manual testing of search endpoint with sample embeddings
• Verify scope and topic filtering work correctly
• Verify similarity scores are in 0-1 range
• Verify limit parameter controls result count
• Test via API Gateway proxy

🤖 Generated with Claude Code