Single tool 'save_memory' has a basic description but lacks critical detail. Input schema is present but minimal, array of strings with no constraints, validation rules, or format guidance. No output schema documented. Description (67 chars) is below the 194-char baseline and does not explain when to use the tool, what happens on duplicates, or error cases. No parameter-level descriptions beyond the array type. Tool lacks idempotency guarantees, error recovery guidance, and security context (who can save memories for whom?). This is a write operation but lacks confirmation/dry-run patterns.

• Add output schema: document what the tool returns (e.g. {saved_count: int, duplicates_detected: int, memory_ids: [string]}). This enables chaining and error detection.
• Expand parameter description: 'memories: array of strings, each 1 - 500 characters, representing facts or context to persist. Exact duplicates are silently skipped; near-duplicates are saved separately.'
• Define duplicate detection: clarify whether duplicates are exact string matches, case-insensitive, or semantic. Return a response field indicating which memories were saved vs. skipped.
• Add error recovery guidance: 'If save fails, the LLM should ask the user to confirm the memory content and retry. Partial saves are not supported, all or nothing.'
• Document user context: 'Memories are saved for the authenticated user making the request. The agent cannot save memories for other users. Audit logs record who saved what and when.'