Loading ToolBench
Mebular MCP server (stdio + Streamable HTTP): exposes Mebular memory capabilities to any MCP client
Mebular exposes 11 memory-management tools with complete JSON schemas and Chinese descriptions. However, descriptions are minimal (10-30 chars), lack LLM-optimized guidance on WHEN to use each tool, and omit parameter-level documentation. Schemas are present but sparse: most parameters lack descriptions, and error handling is absent. Tool naming follows verb_noun convention (memory_*) but lacks clarity on distinctions (e.g., memory_query vs memory_search both retrieve, but criteria differ). No tool annotations (readOnlyHint, destructiveHint). Output schemas are not documented. This is typical of a domain-specific system that prioritizes backend functionality over agent usability.
从起点遍历（深度 ≤ 5）
按会话/时间/主题筛选会话情节
经适配器导入异构来源（kv/markdown/…）
返回用户偏好与属性
语义/关键词召回（配置向量索引时走语义）
关键词检索基线，可选一跳关系
按分类/关键词/标签筛选技能
Descriptions are too short (10 - 30 chars) and lack LLM-optimized guidance. E.g., 'memory_query' is described as '语义/关键词召回（配置向量索引时走语义）' (~30 chars). No explanation of WHEN to use memory_query vs memory_search, what each returns, or prerequisites.
Parameter descriptions are missing or minimal. E.g., 'memory_query' has 'query' (string) with no description of format, length, or expected content. 'types' array lacks explanation of valid enum values or when to filter by type.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2025-06-18+ | v2 |
返回设备/网络/计数/状态哈希/开关
连接对端并等待一次双向同步（network.enabled 需为 true）
写入单条记忆（fact/episode/skill/preference/observation）
批量写入记忆（最多 100 条）
No output schemas documented. Tools return results but LLMs cannot infer response structure. E.g., does memory_query return an array of objects? What fields? Are results paginated? No guidance on chaining (e.g., does memory_query return IDs that memory_write accepts).
No error handling guidance. Tools lack descriptions of failure modes, recovery steps, or actionable error messages. E.g., if memory_sync fails due to network, should the agent retry? Call memory_status first? No guidance.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLMs cannot distinguish safe reads from destructive writes. E.g., memory_write and memory_sync modify state but lack destructiveHint; memory_query is read-only but lacks readOnlyHint.