MCP server that registers memory management tools (memory_add, memory_search, memory_list, memory_get, memory_update, memory_delete, memory_delete_scope, memory_list_scopes, memory_digest, conversation_search) for a dual-system layered memory SDK
Dual-mem MCP server has 10 tools with basic schemas and descriptions, but significant gaps prevent higher scores. All tools have descriptions (10-150 chars, mostly adequate), and input schemas are present with type definitions. However, descriptions lack LLM-optimized guidance on WHEN to use each tool vs. alternatives, parameter relationships are undocumented (e.g., content vs. messages in memory_add), and output schemas are completely absent. Error handling is not visible in tool definitions. Tool naming is clear (verb_noun pattern), but composition issues exist: memory_delete and memory_delete_scope are separate tools where one could subsume the other. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Parameter descriptions are minimal (10-30 chars typical), violating the 72-char baseline. The server lacks idempotency guidance, confirmation patterns for destructive ops, and recovery hints in error cases.
检索 L1 原文(含提取后 SHADOW),用于核对原话/时间/来源。
写入记忆。content 与 messages 二选一;messages 为 [{role, content}, ...] 多轮对话。必填 user_id;app_id 省略则用 default_app_id。
删除单条记忆。
按 scope 批量删除。confirm 必须为 true;app_id 省略则用 default_app_id。
触发 System2 深度记忆巩固(dual 模式)。
按 memory_id 获取单条;不存在返回 null。
列出某 scope 下 ACTIVE 记忆。
Output schemas completely undocumented. No tool returns a documented response structure. LLMs cannot plan downstream calls or extract required fields (e.g., memory_id from memory_add, evolution_chain from memory_search). Violates pattern:tool and pattern:response-shaper.
Parameter descriptions are minimal (10-30 chars vs. 72-char baseline). E.g., 'User identifier' for user_id, 'Memory content text' for content. Lack actionable constraints, format hints, or dependency documentation. Violates pattern:tool-description.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | <=2025-11-25 | v2 |
列出存储中已有的记忆 scope(app_id + user_id + agent_id)。
语义检索,结果按 profile / proactive / normal 三路分组;有演化历史的记忆带 evolution_chain(最新→最旧)。
更新记忆正文并重新 embedding。
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. memory_delete and memory_delete_scope are DESTRUCTIVE; memory_search and memory_list are READ_ONLY. Agents cannot infer safety without annotations. Violates current MCP spec (2026-07-28).
Undocumented parameter relationships. memory_add accepts content OR messages (mutually exclusive), but no description states this. memory_delete_scope requires confirm=true but no guidance on why or what happens if false. Violates pattern:tool-description.
No error handling guidance in tool definitions. No recovery hints (e.g., 'If memory not found, try memory_search'). No categorization of errors as retryable vs. user-fixable. Violates pattern:recovery-guide and pattern:error-classification.