Industrial AI memory base for Agents via MCP: SQLite SoT, rebuildable vectors, progressive recall, defensive tool contracts.
Server has 9 tools with mixed quality. Naming is inconsistent (strata_* prefix for ops, but commit_memory, recall_context lack verb clarity). Descriptions exist but are verbose, mixing Chinese and English, and lack LLM-optimized structure. Schemas are present but incomplete: many parameters lack type definitions (e.g., strata_init.mode is enum but others are underspecified). Critical issue: descriptions exceed 1024 chars and contain operational warnings that should be in error handling, not descriptions. No output schemas documented. Error handling is minimal, no recovery guidance or categorization.
作用:将萃取后的高价值事实持久化到 SQLite Truth Store(经 Quality Kernel 校验)。 触发:跨多轮对话确认的客观事实、用户显式偏好、可复用操作规程;或需暂存的 session scratch。 禁忌:拒绝写入临时情绪、假设推演、密码/Token/API Key;禁止模糊时间(刚才/昨天);严禁把未经确认的推测写成 factual_truth。
Expand memory detail from a specific memory record.
Promote session-scoped memories to durable layers based on usage and confidence.
Recall context from memory layers (L0/L1/L2) based on query.
作用:一致性检查(git-fsck 风格):Truth Store 完整性、向量索引漂移、API Key 健康、秘密残留。 触发:部署后首次诊断、索引重建前、API Key 变更后。 禁忌:禁止在生产高峰期运行(扫描全表);禁止忽视 SECRET_RESIDUAL 错误。
作用:初始化 Memory Palace:写入配置、创建 SQLite Truth Store 与向量目录。 触发:首次使用或重置部署时调用一次。 禁忌:禁止在已有生产数据上重复 init 覆盖配置而不备份;禁止把 API Key 写进对话日志。
Descriptions are verbose (200+ chars), mix Chinese/English, and embed operational warnings ('禁忌') that belong in error handling, not tool descriptions. LLMs cannot parse multi-language text reliably.
Tool names lack action verbs. 'strata_init', 'strata_doctor', 'strata_stats', 'strata_project' are noun-heavy and do not clearly signal what action occurs. Should be 'initialize_memory', 'diagnose_memory', 'get_memory_stats', 'export_memory_projection'.
No output schemas documented. Tools return data but LLMs cannot plan downstream calls or extract fields. E.g., recall_context returns what? A list of memories? A single context object? Fields unknown.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | 2026-07-28+ | v2 |
作用:导出 Truth Store 为 Markdown 投影(只读、Git 可审计)。 触发:备份、审计、人工审查、迁移前快照。 禁忌:禁止编辑投影后写回(投影是只读的);禁止在大规模数据上频繁调用。
作用:从 Truth Store 重建向量索引(幂等)。 触发:索引损坏、向量漂移、API Key 变更后重新嵌入。 禁忌:禁止在索引服务中断时调用;禁止中途杀死进程(会留下孤立向量)。
作用:统计 Truth Store 与向量索引的规模、分布、新鲜度。 触发:定期监控、容量规划、性能基准。 禁忌:禁止用于实时告警(延迟高);禁止在扫描中断后信任结果。
Error handling is absent. No error categorization (retryable vs fatal), no recovery guidance, no invalid-value feedback. ToolError class exists but is not integrated into tool responses.
API key handling in strata_init description warns against writing keys to logs, but this is a tool-level concern, not a user-facing description. Parameter descriptions lack type hints (api_key is string but no format/length guidance).