Recall demonstrates solid tool design with clear naming (verb-noun pattern), comprehensive parameter descriptions, and well-structured schemas. All 7 tools have explicit definitions in src/tools.ts with inputSchema, descriptions, and proper typing. Descriptions average ~150 chars, within the 10-1024 baseline. Parameters are well-documented with constraints (e.g., limit 1-50, min_similarity 0-1). However, output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal; no recovery guidance or actionable error messages visible. Tool composition is strong (each does one thing), but missing idempotency guarantees and confirmation patterns for destructive operations (forget).

• Document output schemas for all tools. Example: 'Returns {id: string, content: string, tags: string[], similarity?: number, created_at: ISO8601}'. This enables LLMs to plan downstream calls and extract fields correctly.
• Add a 'confirm_forget' tool or require a confirmation parameter (e.g., confirmed: true) for the forget tool. Destructive operations should have a safety gate.
• Enhance error messages with recovery guidance. E.g., 'Memory not found (id: abc123). Try recall() to search by content, or list_recent() to browse available memories.'
• Expand 'stats' description to explain the response structure and use case: 'Returns {memory_count: number, total_tokens: number}. Call this to monitor space usage before adding large memories.'
• Document idempotency: 'Calling remember() twice with identical content returns the same memory ID (idempotent). Safe to retry on network failures.'

Inferred effective spec: 2025-06-18+.

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained