Memory System for Cursor using MCP - Provides persistent context awareness for Claude
Server provides 7 tools with complete JSON schemas and descriptions. Naming follows verb_noun convention (save_, retrieve_, update_, delete_, search_, list_, get_). However, descriptions are generic and lack LLM-optimized guidance on WHEN to use each tool vs. alternatives. Parameter descriptions exist but are minimal (avg ~40 chars vs. baseline 72). No output schemas documented. Error handling is absent from visible code. Memory system lacks idempotency guarantees and confirmation patterns for destructive operations (delete_memory). Vector embedding implementation is placeholder-grade, not production-ready.
Delete a memory entry from the persistent memory system
Get statistics about stored memories
List all memories with optional filtering
Retrieve memories from the persistent memory system
Save a memory entry to the persistent memory system
Search memories using vector similarity search
Update an existing memory entry
No output schemas documented. LLMs cannot infer what fields to expect from tool responses (e.g., does retrieve_memory return id, content, category, tags, priority, created_at, updated_at?). This forces agents to guess and risks downstream tool failures.
delete_memory lacks confirmation or dry-run pattern. Destructive operations should support a confirmation step to prevent accidental data loss. No error guidance if memory ID does not exist.
Descriptions are generic and do not explain WHEN to use each tool vs. alternatives. E.g., retrieve_memory vs. search_memories vs. list_memories, when should an LLM pick each? Descriptions lack dependency hints and use-case guidance.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | <=2025-11-25 | v2 |
Vector embedding implementation is deterministic hash-based, not ML-backed. Similarity search quality is poor. Production systems require real embeddings (OpenAI, Anthropic, or local model). Current approach will not cluster semantically similar memories.
No error handling visible in tool implementations. No recovery guidance for LLMs (e.g., 'Memory not found. Try search_memories() with a partial query.' or 'Database connection failed. Retry in 5 seconds.'). Agents cannot self-correct on failures.