Rainman provides 5 well-named, verb-led tools (recall, remember, context, links, status) with clear, actionable descriptions (avg 150+ chars). All tools have complete input schemas with typed properties and descriptions. However, output schemas are not documented, LLMs cannot predict response structure for chaining. Parameter descriptions are strong (e.g., 'Files you're currently working on. Memories about these files (esp. past failures/fixes) are surfaced...'), but some lack format constraints (e.g., 'query' accepts free-form strings with no length/pattern guidance). Error handling is minimal, no recovery guidance or categorization. Security is sound (no secrets in params, read-only tools clearly marked), but no permission gates or audit trail declarations.

• Document output schemas for all 5 tools. Specify field names, types, and structure (e.g., recall returns array of {id, content, category, tags, file_refs, timestamp}). Include pagination fields (total_count, next_cursor) if applicable.
• Add error handling guidance to recall, remember, links. E.g., 'If no results found, try: (1) broader query terms, (2) different category filter, (3) call context() to see available memories.'
• Constrain 'query' parameter: add maxLength (e.g., 500), minLength (e.g., 1), and description hint: 'Search term (1-500 chars). Use keywords like "voting bias", "auth", "migration".'
• Expand 'context' description to 150+ chars. Explain: 'Returns recent and high-importance memories (sorted by recency and relevance). Call at session start for orientation, or when recall() results are sparse.'
• Expand 'status' description. Explain what each statistic means: 'Returns memory count by category (pattern/solution/failure/decision/convention/note), layer (project/global), sentiment (positive/negative/neutral), and top tags. Use to assess memory coverage.'
• Add permission declarations to remember (write:memory) and recall/context/links/status (read:memory). This enables least-privilege agent configs and audit trails.