Strata demonstrates solid fundamentals: all 5 tools have descriptions, proper naming conventions (verb-noun), and tool annotations present. However, critical gaps exist: input/output schemas are not visible in the provided source code, parameter descriptions are absent, and error handling guidance is minimal. The codebase shows good engineering practices (Zod validation, structured logging, TypeScript), but tool definitions lack the LLM-optimization detail required for A-grade quality. Schema visibility is the primary blocker, without seeing actual Zod shapes or JSON Schema definitions, schema scoring is severely constrained.

• Provide complete Zod schema definitions (or equivalent JSON Schema) for all tool inputs and outputs. Include in source code or documentation. Example: show the shape of 'remember' input (content: string, tags?: string[], metadata?: object) and output (id: string, created_at: ISO8601, ...).
• Add parameter descriptions to every input field. For 'recall', document: 'query (string, 1-1000 chars): Natural language search term or memory snippet. Supports partial matches.', 'limit (integer, 1-100, default 20): Maximum results to return.', 'hybrid (boolean, default true): Use semantic + lexical search; false = lexical only.'
• Implement structured error responses with recovery hints. Example: if 'forget' fails with 'Memory not found', return {error: 'not_found', message: 'Memory ID xyz not found.', suggestion: 'Try recall() to find the correct memory ID.'}
• Add a dry-run parameter to 'forget' (e.g., dry_run: boolean, default false) so agents can preview deletion before committing. Or implement a two-step pattern: forget_request → confirm_forget.
• Document output schemas explicitly. For 'recall', specify: returns {memories: [{id, content, tags, similarity_score, created_at}], total_count, next_cursor?}. Include field types and what each represents for downstream tool chaining.

Inferred effective spec: 2025-06-18+.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry