memory-mcp has clear, well-structured tool definitions with complete input schemas and descriptions for all 5 tools. Naming follows verb_noun convention (memory_write, memory_read, memory_search, memory_list, memory_delete). Descriptions are concise (40-80 chars) and explain what each tool does. All parameters have type definitions and descriptions. However, output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal (no recovery guidance, no actionable error messages). No tool annotations (readOnlyHint, destructiveHint) despite clear risk profiles. Parameter descriptions lack format constraints and validation rules. Missing pagination guidance for memory_search despite 'limit' parameter.

• Add formal output schemas to all tools. Document memory_read response as {key: string, tags: string[], created_at: ISO8601, updated_at: ISO8601, content: string}. Document memory_search as {matches: [{score: number, key: string, snippet: string}], total: number}. This enables LLMs to plan chained calls.
• Add tool annotations: memory_write and memory_delete should have destructiveHint=true; memory_read, memory_search, memory_list should have readOnlyHint=true. This signals to agents which tools are safe to retry and which require confirmation.
• Enhance error messages with recovery hints. When memory_read fails: 'No entry found for key "xyz". Try memory_search(query="xyz") to find similar keys, or memory_list() to see all available keys.' This guides agents to self-correct.
• Add parameter constraints to descriptions. For 'key': 'Unique identifier (1 - 256 alphanumeric + underscore/hyphen)'. For 'query': 'Search terms (substring match across key, content, tags)'. For 'limit': 'Max results to return (1 - 100, default 10)'.
• Document idempotency guarantees. memory_write with the same key/content is idempotent (preserves created_at). memory_delete is idempotent (returns success even if key already absent). This helps agents reason about retry safety.