Server has 17 tools with complete input schemas and descriptions, but quality is inconsistent. Naming follows verb_noun pattern (vmem_store, vmem_search, etc.), which is good. However, descriptions are in Chinese, making LLM optimization difficult for English-speaking agents. Parameter descriptions are present but often terse (avg ~40 chars). Output schemas are not documented, responses are JSON strings without declared field structures. Error handling is minimal; most tools return JSON with 'ok' flag but lack recovery guidance. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Destructive tools like vmem_merge and vmem_snapshot_import lack confirmation patterns.
添加因果关系链接。
查询记忆的因果链。
获取记忆的直接因果链接。
基于向量余弦相似度查找重复记忆对。扫描所有记忆的 embedding,计算两两相似度,返回高于阈值的重复对。当记录数超过 10000 时自动启用分块计算以控制内存。
返回记忆系统健康报告。
添加记忆之间的关系链接。
合并两条记忆为一条,保留幸存者并删除另一条。支持三种合并策略:keep_richer(保留信息更丰富的那条)、keep_newer(保留 updated_at 较新的那条)、manual(以 key_a 为幸存者,合并 key_b 的部分信息)。合并后幸存者的置信度提升 10%(上限 1.0),原有的因果链接和记忆链接会被转移。
Descriptions in Chinese; LLM optimization requires English. Descriptions avg 40 chars, below 50-200 char baseline for A+ tools. Terse descriptions like '返回向量记忆数据库的统计信息' (vmem_stats) lack context for tool selection.
No output schemas documented. Tools return JSON strings (e.g., vmem_search returns formatted array, vmem_stats returns dict) but field structures are not declared. LLMs cannot plan downstream calls or extract data reliably.
Destructive tools (vmem_merge, vmem_snapshot_import) lack confirmation or dry-run patterns. vmem_merge with strategy='keep_richer' auto-selects survivor; vmem_snapshot_import with mode='overwrite' clears all data. No recovery guidance in error responses.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
查询用户画像属性。
设置用户画像属性。
检索教训和经验。
查询记忆的关联记忆。
语义搜索记忆条目(融合向量+全文+时间衰减+置信度)。
导出当前记忆数据库的完整快照为 JSON 文件。包含所有 memory_vectors 记录(向量以 base64 编码)、causal_links、memory_links、topics、session_topics 等关联表数据,以及导出元信息。
从 JSON 快照文件导入记忆数据。支持三种导入模式:merge(已有 key 跳过,只导入新记录)、overwrite(清空当前数据后全量导入)、upsert(对比 updated_at,保留较新的记录)。导入后会自动重建 FTS5 全文索引并运行 schema 迁移以确保兼容。
返回向量记忆数据库的统计信息。
存储一条记忆到向量数据库。
列出所有话题聚类。
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk levels are known (READ_ONLY, WRITE, DESTRUCTIVE) but not declared in tool metadata. Agents cannot reason about safety or retry semantics.
Error handling minimal. Most tools return {'ok': False, 'error': str(e)} with raw exception text. No recovery guidance, categorization (retryable vs fatal), or actionable next steps for LLM.