Engram demonstrates solid tool design with clear naming (verb_noun pattern), comprehensive parameter descriptions, and well-structured schemas using Zod. All 7 tools have descriptions (avg ~150 chars) and typed parameters with constraints. However, output schemas are not formally documented, responses are text-only without structured field definitions. Error handling is minimal (no recovery guidance or categorization). The semantic memory domain is well-served, but the tool set lacks composition patterns (e.g., batch operations) and some parameter relationships are underdocumented.

• Document output schemas formally. Define structured responses with typed fields (e.g., { memories: [{ id: number, content: string, type: string, similarity?: number }], total: number }) instead of text-only results.
• Add error categorization and recovery guidance. E.g., 'Memory #999 not found. Call list_memories() to see available IDs.' Distinguish retryable errors from user-fixable ones.
• Implement confirmation pattern for destructive operations. Add a 'dry_run' parameter to forget and update_memory, or require explicit confirmation via a separate confirm_operation tool.
• Document parameter dependencies in descriptions. Clarify how 'type' and 'tags' affect deduplication in save_memory, and explain when to use find_related before save_memory to avoid duplicates.
• Add pagination support to recall and find_related. Include cursor/offset and total_count in responses to handle large result sets without blowing context windows.