mem-mesh has 9 tools with complete input schemas and descriptions. Naming is verb-forward (search_, get_, list_, add_, batch_*) and mostly clear. Descriptions are present and actionable (avg ~120 chars), though some lack WHEN-to-use guidance. Parameters are well-typed with constraints (enums, min/max, nullable flags). Output schemas are not explicitly documented in the source, the rubric requires visible schema definitions for full credit. Error handling is minimal; no recovery guidance or categorization visible. Security: no credentials in params (good), but no audit logging or permission gates evident. Composition is strong: tools chain well (search → get_memory_context), batch variants exist (batch_add_memories, batch_search), and idempotency is implicit for read tools.

• Document output schemas for all tools. Include field names, types, and pagination info. Example: 'Returns {memories: [{id, content, category, created_at, score}], total: int, next_cursor: string|null}'.
• Add error handling with recovery guidance. E.g., 'If search returns 0 results, try: (1) broaden query, (2) remove category filter, (3) call memory_stats to verify data exists.'
• Rename 'add' to 'add_memory' for consistency and clarity. Update description to explain when to use add vs batch_add_memories (single vs bulk).
• Enhance descriptions with WHEN-to-use hints. E.g., 'Use search_memories for keyword/vector queries; use memory_stats to see aggregate counts before searching.'
• Add permission checks and audit logging to destructive tools. Implement dry-run support for add and batch_add_memories. Log caller, timestamp, and operation outcome.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry