A Model Context Protocol (MCP) server for managing user memories with web interface
Server has 6 tools with complete input schemas and descriptions, but quality is inconsistent. Tool names follow verb_noun convention (store_memory, list_memory_titles, get_memory_by_title, delete_memory, open_memory_web, stop_memory_web). Descriptions are present but lack actionable guidance for LLM selection. Critical gaps: no output schemas documented, no error recovery guidance, no parameter validation rules stated, descriptions in Chinese limit LLM comprehension, and no tool annotations (readOnlyHint/destructiveHint). The open_memory_web and stop_memory_web tools have problematic design, they manage side effects (browser launch, process termination) without confirmation or dry-run patterns, violating the confirmation-request pattern for irreversible operations.
删除指定标题的记忆。
根据用户选择的标题获取记忆的完整内容。
获取所有记忆的标题列表,用于在聊天窗口中展示供用户选择。只返回标题和基本信息,不返回完整内容。
打开记忆管理Web界面。当用户说出以下任一触发词时自动调用:'打开记忆MCP'、'打开记忆'、'打开记忆WEB'、'开启记忆'。会自动启动Web服务器并在浏览器中打开管理界面。
停止记忆管理Web界面服务器。用于强制停止所有相关进程和释放端口。
存储新的记忆。AI助手必须先要求用户提供记忆标题,然后要求用户输入记忆内容。
No output schemas documented. Tools return success/error objects but LLMs cannot plan downstream calls without knowing response structure (e.g., what fields does executeStoreMemory return? Is 'id' always present?). This violates the response-shaper pattern.
Descriptions are in Chinese, limiting comprehension by English-speaking LLMs and reducing tool selection accuracy. Descriptions should be in the LLM's primary language (typically English for production systems).
open_memory_web and stop_memory_web lack confirmation/dry-run patterns. These tools trigger irreversible side effects (launching browsers, killing processes) without user confirmation. Agents may invoke them unintentionally, violating the confirmation-request pattern.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | 2026-07-28+ | v2 |
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). delete_memory is destructive but unmarked; list_memory_titles and get_memory_by_title are read-only but unmarked. Annotations help LLMs reason about safety and retry logic.
Error responses lack recovery guidance. executeStoreMemory returns {success: false, error: '...'} but does not tell the LLM what to do next (retry? ask user? abort?). Violates the recovery-guide pattern.