Loading ToolBench
Zero-knowledge persistent memory layer for Claude — MCP server with AES-256-GCM encryption
MemoraEU MCP has 5 tools with French descriptions and reasonable naming (remember, recall, forget, list_memories, update_memory). However, critical gaps undermine quality: (1) Input schemas are partially visible but incomplete, parameters lack explicit type constraints and enums; (2) Descriptions are action-oriented but lack LLM-critical details like when to use each tool, prerequisites, and error recovery guidance; (3) No documented output schemas, LLMs cannot infer what fields to expect from recall or list_memories; (4) Parameter descriptions are minimal (e.g., 'Identifiant unique' for id) and do not specify format, constraints, or valid ranges; (5) No error handling guidance, tools do not indicate what to do on failure (e.g., 'User not found. Try search_users()'); (6) Security concerns: MEMORAEU_API_KEY and MEMORAEU_SECRET are environment variables but no explicit secret-injection pattern documented in tool definitions; (7) Composition: tools are single-responsibility (good), but output chaining is unclear, does recall return memory IDs that forget accepts? Not documented.
Supprime une mémoire spécifique par son identifiant.
Liste toutes les mémoires de l'utilisateur avec pagination et filtrage optionnel par catégorie.
Recherche les mémoires pertinentes par requête textuelle avec recherche sémantique. Injecte automatiquement le contexte de session au premier appel.
Mémorise une information importante avec chiffrement zero-knowledge. Détecte automatiquement les doublons, compresse si nécessaire, et suggère une catégorie.
Met à jour le contenu ou la catégorie d'une mémoire existante avec re-chiffrement.
No documented output schemas. LLMs cannot infer what fields recall(), list_memories(), or remember() return. This forces agents to guess at response structure and breaks downstream tool chaining.
Parameter descriptions are minimal and lack constraints. E.g., 'limit' has no min/max bounds (should be 1 - 100 or similar); 'category' has no length limit or character restrictions; 'id' does not specify format (UUID? alphanumeric?). LLMs cannot validate inputs without explicit constraints.
No error handling guidance. Tools do not indicate what to do on failure (e.g., 'Memory not found' for forget/update_memory, 'API timeout' for recall). Agents have no recovery path.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
Tool descriptions lack LLM-critical context. No explanation of WHEN to use each tool (e.g., 'Use recall to retrieve relevant memories before responding to user queries'). No prerequisites or dependencies documented. Descriptions are action-only, not decision-guiding.
Destructive operations (forget, update_memory) lack confirmation or dry-run support. Agents can permanently delete memories without a safety gate. No idempotency guarantees documented.