Loading ToolBench
Fragment-Based Memory MCP Server with multi-round-trip request support, LLM fallback chains, morpheme-based embedding, and symbolic rule evaluation
Memento MCP has 20 tools with inconsistent quality. While all tools have basic descriptions and input schemas are registered, the definitions lack critical LLM-optimization details required for production agentic use. Naming is generally acceptable (verb-based: remember, recall, forget, etc.), but descriptions are terse (most 10-50 chars) and parameters lack sufficient constraint documentation. Many parameters missing descriptions entirely (e.g., 'remember' tool only documents topic/type but no parameter descriptions visible in schema). No documented output schemas. Error handling guidance is absent. Security considerations (master key requirements on tools 19-20) are mentioned in registry metadata but not in descriptions. Critical gap: tool descriptions do not explain WHEN to use each tool vs similar ones (recall vs graph_explore vs reconstruct_history are semantically overlapping). Per-tool scores average 42 across 20 tools.
Update or correct an existing memory fragment
指定されたステップの更新を実行します。ユーザー同意後、AI が呼び出し。master key 専用。
Store multiple memory fragments in a single batch operation
Retrieve the status of a batch memory operation
現在のバージョンと最新GitHub タグを比較して更新可能性を確認します。master key 専用。
Load contextual memory fragments for the current conversation
Delete a memory fragment by ID or topic
Descriptions frequently under 50 characters and lack LLM-optimization context (when to use each tool, what output is returned, how this tool differs from similar ones)
Parameter-level descriptions missing or absent from visible schema definitions. The 'remember' tool schema shows properties but no 'description' fields for topic/type parameters.
No output/response schemas documented for any tool. LLMs cannot plan downstream chaining without knowing what fields are returned.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 53 | <=2025-11-25 | v2 |
| 2026-03-09 | C | 61 | 2025-11-25+ | v1 |
Retrieve the edit and version history of a memory fragment
Retrieve skill guide and usage documentation
Explore memory relationships and traces from a starting fragment
Create a relationship link between two memory fragments
Execute memory consolidation and garbage collection (admin only)
Retrieve memory statistics and analytics (admin only)
Retrieve memory fragments using keywords, topic, or hybrid search
Reconstruct historical memory traces for a case or entity
Analyze and consolidate memories from a session
Store a memory fragment with topic, type, and content
Search memory traces by keyword and case ID
Rotate and reset the current session context
Provide feedback on tool relevance and sufficiency
Semantic overlap and lack of differentiation between similar tools: recall vs graph_explore vs reconstruct_history all retrieve memory; fragment_history vs reconstruct_history both provide history; no description clarifies when to use each.
Descriptions of destructive/sensitive tools (forget, memory_consolidate, apply_update) do not adequately warn of irreversibility or safety implications; destructive metadata is in registry but not visible in tool description presented to LLM.
Non-English descriptions (Japanese) in check_update and apply_update violate accessibility and LLM-optimization baselines; non-English content breaks schema parsing and reduces utility for international teams.
Parameter constraints and enums not documented: recall's 'keywords' can be string|array with no guidance on array length or max keywords; link has no relationship type enum; reconstruct_history has ambiguous required parameters (caseId vs entity).
No error handling guidance visible. Tools return structured success/failure but descriptions do not explain how LLM should respond to errors (retry vs ask user vs fallback).
Idempotency not documented. Registry marks some tools idempotent (batch_status, recall, context) but descriptions do not state this; agents cannot plan retry behavior without this knowledge.