Memory Garden presents a specialized cognitive tool suite with clear, domain-specific descriptions in Chinese. All 9 tools have explicit descriptions and input schemas with type definitions. However, several tools lack output schema documentation, parameter descriptions are minimal (averaging ~40 chars vs. baseline 72), and error handling is inconsistent. Tool naming follows verb_noun convention well (search_, read_, get_, find_, discover_, ask_). The server implements a thoughtful 'boundary' pattern (appending constraints to observations), but this is custom and not MCP-standard. No tool annotations (readOnlyHint, destructiveHint) despite all tools being read-only. Parameter validation exists but error messages are sparse.

• Document output schemas for all tools. Specify the structure of the 'data' dict returned by each tool (e.g., search_sources returns {query, hits: [{atom_id, title, excerpt, ...}]}). Include field types and whether fields are always present.
• Expand parameter descriptions to 50-150 characters with actionable constraints. E.g., 'limit: Maximum number of results to return (1-20, default 6). Larger limits may slow retrieval.' Include format hints for date parameters (ISO 8601).
• Add readOnlyHint annotation to all 9 tools in the MCP tool registration. This signals to clients that these tools are safe to cache and parallelize.
• Implement structured error responses with recovery guidance. E.g., if atom_id is not in _seen_atom_ids, return {error: '...', recovery: 'Call search_sources() first to discover sources.'}
• Decompose ask_garden into smaller, composable tools. Instead of a meta-agent, expose the underlying retrieval and reasoning steps (e.g., retrieve_evidence, synthesize_conclusion) so agents can control the flow and inspect intermediate results.