Loading ToolBench
MCP server exposing the English Memory Agent's SQLite storage as tools. Lets external MCP clients (Claude Desktop, Cursor, ...) save, search, list, delete, and quiz over memory cards without going through the ADK workflow. Card saves still pass through the same privacy_scan gate as the agent path.
Server has 5 tools with complete input schemas and descriptions, but quality is uneven. Tool names follow verb_noun convention (save_, search_, list_, delete_, generate_). Descriptions are present (avg ~100 chars) but lack actionable context for LLM selection. Parameter descriptions exist but are often generic ('The search keyword'). No output schemas documented. Error handling is minimal, tools return status/message dicts but lack recovery guidance or categorization. save_memory_card has 12 optional parameters with weak descriptions ('Detailed Chinese explanation and usage notes' is vague). No enum constraints on card_type despite it accepting only 4 values. Privacy scanning is implemented but not surfaced in tool descriptions. Overall: functional but below production baseline.
Deletes a memory card from the local SQLite database by its ID.
Generates a list of review quiz questions based on the most recently saved cards.
Lists the most recently saved memory cards in the database.
Saves an English vocabulary or error card into the local SQLite memory database.
Searches the memory database for saved cards matching the keyword. Looks across original sentences, corrections, Chinese explanations, key expressions, scenarios, and tags.
save_memory_card lacks enum constraint on card_type parameter. Description says 'Must be error_card, rewrite_card, expression_card, or word_card' but no enum in schema. LLMs will hallucinate invalid values.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract structured data. E.g., search_memory_cards returns 'list[dict]' but field names/types are unknown.
Error responses lack recovery guidance. delete_memory_card returns {'status': 'error', 'message': 'Card ID X not found'} but does not suggest next steps (e.g., 'Try list_recent_memory_cards to find valid IDs').
save_memory_card description does not mention privacy scanning or that inputs are validated for PII. This is a critical feature that should be surfaced to the LLM.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
Parameter descriptions are generic. E.g., 'The search keyword (in English or Chinese)' lacks guidance on multi-word queries, partial matches, or expected result count. LLMs cannot optimize queries.