Memory-augmented reasoning for AI agents - An MCP server that enables persistent memory management for AI agents with semantic deduplication, intelligent memory extraction, and retrieval capabilities.
ReasoningBank MCP provides two tools with explicit schema definitions and descriptions. However, the definitions exhibit significant gaps in clarity, parameter completeness, and error handling guidance. Tool names follow the verb_noun pattern (retrieve_memory, extract_memory) which is positive. Descriptions are present but written in Chinese and somewhat generic. Schemas are present with proper JSON structure, but lack critical details like min/max constraints, format specifications, and output documentation. The trajectory parameter in extract_memory is particularly complex but under-documented regarding expected structure and validation. No error handling guidance, recovery hints, or security annotations are evident.
从任务轨迹中提取推理经验并保存到记忆库。当任务执行完成后必须调用此工具,以便将经验保存供未来使用。这应该在每次任务结束时调用。
检索相关的历史经验记忆,帮助指导当前任务的执行。在开始执行任何复杂任务前都应该先调用此工具来获取相关经验。例如:编写代码、数据分析、制定计划等任务前。
Missing output schema documentation. Neither tool documents what fields are returned to the LLM, making it impossible for agents to plan downstream steps or extract specific data.
Insufficient parameter validation guidance. top_k in retrieve_memory has no documented min/max bounds (could be 0, 1000, or negative). Default is 1 but no upper limit specified.
extract_memory trajectory parameter is complex (array of objects with nested metadata) but lacks detailed format specification. What fields are required vs optional in metadata? What role enum values map to in real conversation flows? This ambiguity will cause agent errors.
No error handling or recovery guidance. If retrieve_memory finds no matches, should the agent retry with a different query? If extract_memory fails to save, is it retryable? No guidance provided.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 50 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 49 | - | v1 |
Descriptions are in Chinese (non-English). While technically acceptable, this creates barriers for multi-lingual LLM routing and reduces clarity for many agent developers. Should be provided in English as the primary language.
success_signal parameter in extract_memory is optional and defaults to 'null' with automatic judgment. No description of how success is judged, what constitutes success, or how this affects memory storage/retrieval. Undocumented dependency.
async_mode parameter in extract_memory defaults to true but offers no guidance on when to use async vs sync, what the behavioral differences are, or how this affects tool outcome reporting.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). This is a current MCP spec pattern (2026-07-28). retrieve_memory should be marked readOnly. extract_memory should be marked destructive (writes to persistent storage).