memocall exposes 5 well-named, read-only tools for recalling past Claude Code conversations. All tools have clear, actionable descriptions (120-250 chars each) that explain WHAT they do and WHEN to use them. Input schemas are present and use Zod with type definitions and descriptions. However, output schemas are NOT documented, the code returns text responses wrapped in a generic {content: [{type: 'text', text: s}]} structure, but there is no declared schema showing what fields or structure the user should expect. Descriptions are good but lack some dependency hints and error-recovery guidance. Parameter descriptions are solid (average ~50 chars), though some could be more explicit about constraints. All tools follow the verb_noun naming pattern correctly. The tools are read-only (no destructive operations) and properly composed, each has a single responsibility and produces output that chains into other tools (e.g., list_sessions and search_sessions return IDs that load_session accepts). Error handling returns friendly messages but does not categorize errors or offer recovery paths. No tool has examples in descriptions (good). Overall: strong naming, good descriptions, present but undocumented output schemas, and solid parameter coverage. Lacks error recovery guidance and explicit output schema documentation.

• Document output schemas for all tools. For list_sessions, define: {sessions: [{id: string, title: string, date: ISO8601, turnCount: number, cwd: string}], total: number}. For load_session, define: {transcript: string, tokenCount: number, turnRange: [number, number], warnings?: string[]}. This enables LLMs to plan chaining calls.
• Add error recovery guidance. Example: 'No session found. Try search_sessions(query) to find by topic, or list_sessions(limit=100) to browse recent sessions.' This guides the agent on next steps.
• Enforce mutual exclusivity in load_session between 'id' and 'query'. Use Zod .refine() to validate that exactly one is provided, and update the description to state 'Exactly one of id or query is required.'
• Add validation for 'turns' parameter syntax. Use a regex pattern in the description: 'Format: "300-340" (inclusive range), "300-" (from 300 onward), "-50" (last 50 turns), or "300" (single turn). Turn numbers are 1-based.' Consider a Zod .regex() constraint.
• Expand pagination guidance. For list_sessions and search_sessions, document: 'Results are truncated at the limit. To see more sessions, increase limit or use search_sessions(query) to narrow the scope.' Consider returning a 'hasMore' boolean or 'total' count in the output.
• Add dependency hints in descriptions. E.g., list_sessions: 'Start here to see available sessions. Once you have a session ID, call load_session(id) to pull its context.' This prevents discovery detours.
• Categorize errors in implementation. For 'No session found', return: {error: 'NOT_FOUND', message: 'Session not found for query "..."', suggestion: 'Try search_sessions(query) to find by topic.'}. This lets the agent decide to retry, ask the user, or pivot.