Mnemos provides three read-only knowledge-base tools with clear verb-noun naming (search_context, list_documents, get_document_info) and functional JSON schemas. Descriptions are adequate (100-150 chars) but lack strategic depth, they state WHAT the tool does without explaining WHEN to use it or how it chains with other tools. Parameter schemas are present with types and defaults, but descriptions are minimal or absent for some params (e.g., 'collection' param in search_context lacks a description explaining its semantics or how to discover valid values). Output schemas are not documented, LLMs cannot infer what fields to expect from these calls. Error handling is absent from tool descriptions, so agents cannot recover from common failures. The tools are well-composed (each does one thing) and read-only (no destructive operations), but they fall short of production-grade quality due to incomplete parameter documentation, missing output schema specifications, and lack of error recovery guidance.

• Document output schemas for all three tools. Add explicit sections like: 'Returns: [{ document_id: string, title: string, content: string, metadata: {...} }]' to tool descriptions or as structured output documentation.
• Add descriptions to all parameters, even optional ones. For 'collection', explain: 'Filter results to a specific knowledge collection (e.g., "frontend-docs", "api-reference"). Omit to search all collections. Call list_documents with no args to discover available collection names.'
• Extend search_context description to clarify ranking and result guarantees: 'Returns the top-k most relevant document chunks ranked by semantic similarity. Fewer than k results may be returned if fewer match the query.'
• Add error recovery hints to descriptions. E.g. 'If get_document_info returns an error, the document may have been deleted. Call list_documents to verify it exists before retrying.'
• Implement and document pagination for list_documents. Add offset/page parameters, return a total_count field, and clarify the response limit in the description: 'Returns up to [limit] documents. For large collections, use offset to paginate.'
• Consider adding tool annotations (toolAnnotations capability) to mark search_context and list_documents as readOnlyHint: true, signaling that these are safe to call speculatively.