MASE Memory MCP exposes 13 tools with mixed quality. Naming is generally verb-forward (get_, list_, mase_*), but descriptions vary widely: some are detailed (mase_remember, mase_upsert_fact), others are terse or Chinese-only (get_current_time: 'Returns the current system time'). Input schemas are present for all tools with proper type declarations, but parameter descriptions are inconsistent, some parameters lack descriptions entirely (e.g., 'dirpath' in list_directory has a description, but 'filepath' in read_local_file is minimal). Output schemas are not documented. Error handling is mentioned in read_local_file ('Returns a clearly-tagged error string') but absent from most other tools. No tool annotations (readOnlyHint, destructiveHint) are visible. The server mixes English and Chinese descriptions, which may confuse LLM parsing. Overall, tools are functional but lack the polish expected of production-grade agent tools.

• Document output schemas for all tools. For mase_recall, specify: {results: [{log_id: int, text: string, timestamp: string, role: string}], total_count: int}. For mase_get_facts, specify: {facts: [{category: string, key: string, value: string, updated_at: string}]}.
• Standardize descriptions to English only. Provide Chinese translations in separate documentation, not inline. This ensures LLM parsing consistency.
• Add tool annotations: mark mase_remember, mase_upsert_fact, mase_ask_tool as destructiveHint=true; mark all read tools as readOnlyHint=true.
• Clarify mase_ask_tool's purpose: 'Use this only when you need the full MASE reasoning pipeline (routing, memory, planning, execution). For simple fact lookup, use mase_get_facts or mase_recall instead.'
• Add error handling guidance to mase_upsert_fact: 'If category or key is invalid, returns error with valid category list. If source_log_id does not exist, returns warning but proceeds.'